iv

For my wife Tatiana and my son Eugene

vi

Contents
Preface Introduction 1 Basic de nitions and auxiliary statements
1.1 Sets, functions, real numbers . . . . . . . . . . . . . . . . . . . . . 1.1.1 Notations and de nitions . . . . . . . . . . . . . . . . . . . 1.1.2 Real numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Topological, metric, and normed spaces . . . . . . . . . . . . . . . 1.2.1 General notions . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.2 Metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.3 Normed vector spaces . . . . . . . . . . . . . . . . . . . . . 1.3 Continuous functions and compact spaces . . . . . . . . . . . . . . 1.3.1 Continuous and semicontinuous mappings . . . . . . . . . . 1.3.2 Compact spaces . . . . . . . . . . . . . . . . . . . . . . . . 1.3.3 Continuous functions on compact spaces . . . . . . . . . . . 1.4 Maximum function and its properties . . . . . . . . . . . . . . . . . 1.4.1 Discrete maximum function . . . . . . . . . . . . . . . . . . 1.4.2 General maximum function . . . . . . . . . . . . . . . . . . 1.5 Hilbert space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.1 Basic de nitions and properties . . . . . . . . . . . . . . . . 1.5.2 Compact and selfadjoint operators in a Hilbert space . . . . 1.5.3 Theorem on continuity of a spectrum . . . . . . . . . . . . 1.5.4 Imbedding of a Hilbert space in its dual . . . . . . . . . . . 1.5.5 Scales of Hilbert spaces and compact imbedding . . . . . . 1.6 Functional spaces that are used in the investigation of boundary value and optimal control problems . . . . . . . . . . . . . . . . . . 1.6.1 Spaces of continuously di erentiable functions . . . . . . . . 1.6.2 Spaces of integrable functions . . . . . . . . . . . . . . . . . 1.6.3 Test and generalized functions . . . . . . . . . . . . . . . . 1.6.4 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . 1.7 Inequalities of coerciveness . . . . . . . . . . . . . . . . . . . . . . . vii

xv xix 1

1 1 2 4 4 5 6 10 10 13 14 15 15 16 18 18 20 25 31 33

36 36 37 38 39 45

viii

CONTENTS 1.7.1 Coercive systems of operators . . . . . . . . . . . . . . . . . 1.7.2 Korn's inequality . . . . . . . . . . . . . . . . . . . . . . . . Theorem on the continuity of solutions of functional equations . . Di erentiation in Banach spaces and the implicit function theorem 1.9.1 Frechet derivative and its properties . . . . . . . . . . . . . 1.9.2 Implicit function . . . . . . . . . . . . . . . . . . . . . . . . 1.9.3 The G^teaux derivative and its connection with the Frechet a derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . Di erentiation of the norm in the space Wpm ( ) . . . . . . . . . . . 1.10.1 Auxiliary statement . . . . . . . . . . . . . . . . . . . . . . 1.10.2 Theorem on di erentiability . . . . . . . . . . . . . . . . . Di erentiation of eigenvalues . . . . . . . . . . . . . . . . . . . . . 1.11.1 The eigenvalue problem . . . . . . . . . . . . . . . . . . . . 1.11.2 Di erentiation of an operator-valued function . . . . . . . . 1.11.3 Eigenspaces and projections . . . . . . . . . . . . . . . . . . 1.11.4 Di erentiation of eigenvalues . . . . . . . . . . . . . . . . . The Lagrange principle in smooth extremum problems . . . . . . . G-convergence and G-closedness of linear operators . . . . . . . . . Di eomorphisms and invariance of Sobolev spaces with respect to di eomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.14.1 Di eomorphisms and the relations between the derivatives . 1.14.2 Sequential Frechet derivatives and partial derivatives of a composite function . . . . . . . . . . . . . . . . . . . . . . . 1.14.3 Theorem on the invariance of Sobolev spaces . . . . . . . . 1.14.4 Transformation of derivatives under the change of variables 45 48 51 52 52 53 54 55 55 56 59 59 61 62 65 71 73 74 74 76 77 79

1.8 1.9

1.10 1.11

1.12 1.13 1.14

2 Optimal control by coe cients in elliptic systems

2.1 Direct problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Coercive forms and operators . . . . . . . . . . . . . . . . . 2.1.2 Boundary value problem . . . . . . . . . . . . . . . . . . . . 2.2 Optimal control problem . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Nonregular control . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Regular control . . . . . . . . . . . . . . . . . . . . . . . . . 2.2.3 Regular problem and necessary conditions of optimality . . 2.2.4 Nonsmooth (discontinuous) control . . . . . . . . . . . . . . 2.2.5 Some remarks on the use of regular and discontinuous controls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 The nite-dimensional problem . . . . . . . . . . . . . . . . . . . . 2.4 The nite-dimensional problem (another approach) . . . . . . . . . 2.4.1 The set U (t) . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.2 Approximate solution of the problem (2.2.22) . . . . . . . . 2.4.3 Approximate solution of the optimal control problem when the set Uad is empty . . . . . . . . . . . . . . . . . . . . . . 2.4.4 On the computation of the functional h ! k (h uh ) . . . .

81

81 81 83 86 86 88 90 97

102 103 106 106 107 109 111

CONTENTS

ix

2.5 2.6

2.7

2.8 2.9 2.10 2.11

2.12

2.13

2.4.5 Calculation and use of the Frechet derivative of the functional h ! m (h uh ) . . . . . . . . . . . . . . . . . . . . . 113 Spectral problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 2.5.1 Eigenvalue problem . . . . . . . . . . . . . . . . . . . . . . 118 2.5.2 On the continuity of the spectrum . . . . . . . . . . . . . . 118 Optimization of the spectrum . . . . . . . . . . . . . . . . . . . . . 120 2.6.1 Formulation of the problem and the existence theorem . . . 120 2.6.2 Finite-dimensional approximation of the optimal control problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 2.6.3 Computation of eigenvalues . . . . . . . . . . . . . . . . . . 128 Control under restrictions on the spectrum . . . . . . . . . . . . . 130 2.7.1 Optimal control problem . . . . . . . . . . . . . . . . . . . . 130 2.7.2 Approximate solution of the problem (2.7.7) . . . . . . . . . 131 2.7.3 Second method of approximate solution of the problem (2.7.7) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132 2.7.4 Di erentiation of the functionals h ! Ai (h) and necessary conditions of optimality . . . . . . . . . . . . . . . . . . . . 136 The basic optimal control problem . . . . . . . . . . . . . . . . . . 139 2.8.1 Setting of the problem. Existence theorem . . . . . . . . . . 139 2.8.2 Approximate solution of the problem (2.8.6) . . . . . . . . . 140 The combined problem . . . . . . . . . . . . . . . . . . . . . . . . . 143 Optimal control problem for the case when the state of the system is characterized by a set of functions . . . . . . . . . . . . . . . . . 145 2.10.1 Setting of the problem . . . . . . . . . . . . . . . . . . . . . 145 2.10.2 The existence theorem . . . . . . . . . . . . . . . . . . . . . 147 The general control problem . . . . . . . . . . . . . . . . . . . . . . 150 2.11.1 Bilinear form aq and the corresponding equation . . . . . . 150 2.11.2 Bilinear form br and the spectral problem . . . . . . . . . . 153 2.11.3 Basic control problem . . . . . . . . . . . . . . . . . . . . . 154 2.11.4 Application of the basic control problem (combined problem) 158 Optimization by the shape of domain and by operators . . . . . . . 160 2.12.1 Domains and bilinear forms . . . . . . . . . . . . . . . . . . 160 2.12.2 Optimization problem connected with solution of an operator equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 2.12.3 Eigenvalue optimization problem . . . . . . . . . . . . . . . 163 2.12.4 Some realizations of the spaces Ml and Nl . . . . . . . . . . 164 Optimization problems with smooth solutions of state equations . 168 2.13.1 Systems of elliptic equations . . . . . . . . . . . . . . . . . . 168 2.13.2 Elliptic problems in domains and in a xed domain . . . . . 171 2.13.3 The problem of domain shape optimization . . . . . . . . . 174 2.13.4 Approximate solution of the direct problem ensuring convergence in the norm of a space of smooth functions . . . . 175

CONTENTS 3.1 On the minimum of nonlinear functionals . . . . . . . . . . . . . . 3.1.1 Setting of the problem. Auxiliary statements . . . . . . . . 3.1.2 The existence theorem . . . . . . . . . . . . . . . . . . . . . 3.1.3 Characterization of a minimizing element . . . . . . . . . . 3.1.4 Functionals continuous in the weak topology . . . . . . . . 3.2 Approximate solution of the minimization problem . . . . . . . . . 3.2.1 Inner point lemma . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Finite-dimensional problem . . . . . . . . . . . . . . . . . . 3.3 Control by the right hand side in elliptic problems provided the goal functional is quadratic . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 Setting of the problem . . . . . . . . . . . . . . . . . . . . . 3.3.2 Existence of a solution. Optimality conditions . . . . . . . . 3.3.3 An example of a system described by the Dirichlet problem 3.4 Minimax control problems . . . . . . . . . . . . . . . . . . . . . . . 3.5 Control of systems whose state is described by variational inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5.1 Setting of the problem . . . . . . . . . . . . . . . . . . . . . 3.5.2 The existence theorem . . . . . . . . . . . . . . . . . . . . . 3.5.3 An example of control of a system described by a variational inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3 Control by the right hand sides in elliptic problems

179

179 179 181 183 184 185 185 187

193 193 194 196 201 204 204 205 208

4 Direct problems for plates and shells

4.1 Bending and free oscillations of thin plates . . . . . . . . . . . . . . 4.1.1 Basic relations of the theory of bending of thin plates . . . 4.1.2 Orthotropic plates . . . . . . . . . . . . . . . . . . . . . . . 4.1.3 Bilinear form corresponding to the strain energy of the plate 4.1.4 Problem of bending of a plate . . . . . . . . . . . . . . . . . 4.1.5 Problem of free oscillations of a plate . . . . . . . . . . . . 4.2 Problem of stability of a thin plate . . . . . . . . . . . . . . . . . . 4.2.1 Stored energy of a plate . . . . . . . . . . . . . . . . . . . . 4.2.2 Conditions of stationarity . . . . . . . . . . . . . . . . . . . 4.2.3 Auxiliary statements . . . . . . . . . . . . . . . . . . . . . . 4.2.4 Transformation of the problem (4.2.27), (4.2.28) . . . . . . 4.2.5 Stability of a plate and bifurcation . . . . . . . . . . . . . . 4.2.6 An example of nonexistence of stable solutions . . . . . . . 4.3 Model of the three-layered plate ignoring shears in the middle layer 4.3.1 Basic relations . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2 Problems of the bending and of the free exural oscillations 4.4 Model of the three-layered plate accounting for shears in the middle layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.1 Basic relations . . . . . . . . . . . . . . . . . . . . . . . . . 4.4.2 Bilinear form corresponding to the three-layered plate . . . 4.4.3 Bending of the three-layered plate . . . . . . . . . . . . . .

211

211 211 213 214 217 223 225 225 228 230 233 237 242 244 244 246

248 248 253 255

CONTENTS 4.4.4 Natural oscillations of three-layered plate . . . . . . . . . . 4.5 Basic relations of the shell theory . . . . . . . . . . . . . . . . . . . 4.6 Shells of revolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6.1 Deformations and functional spaces . . . . . . . . . . . . . 4.6.2 The bilinear form ah . . . . . . . . . . . . . . . . . . . . . . 4.6.3 The subspace of functions with zero-point strain energy . . 4.7 Shallow shells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8 Problems of statics of shells . . . . . . . . . . . . . . . . . . . . . . 4.9 Free oscillations of a shell . . . . . . . . . . . . . . . . . . . . . . . 4.10 Problem of shell stability . . . . . . . . . . . . . . . . . . . . . . . 4.10.1 On some approaches to stability problems . . . . . . . . . . 4.10.2 Reducing of the stability problem to the eigenvalue problem 4.10.3 Spectral problem (4.10.12) . . . . . . . . . . . . . . . . . . . 4.11 Finite shear model of a shell . . . . . . . . . . . . . . . . . . . . . . 4.11.1 Strain energy of an elastic shell . . . . . . . . . . . . . . . . 4.11.2 Shallow shell . . . . . . . . . . . . . . . . . . . . . . . . . . 4.11.3 A relation between the Kirchho and Timoshenko models of shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.12 Laminated shells . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.12.1 The strain energy of a laminated shell . . . . . . . . . . . . 4.12.2 Shell of revolution . . . . . . . . . . . . . . . . . . . . . . . 4.12.3 Shallow shells . . . . . . . . . . . . . . . . . . . . . . . . . .

xi 258 260 263 263 265 267 268 270 272 273 273 274 275 277 277 279 281 285 285 287 289

5 Optimization of deformable solids

5.1 Settings of optimization problems for plates and shells . . . . . . . 5.1.1 Goal functional and a function of control . . . . . . . . . . 5.1.2 Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Approximate solution of direct and optimization problems for plates and shells . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1 Direct problems and spline functions . . . . . . . . . . . . . 5.2.2 The spaces Vm for plates . . . . . . . . . . . . . . . . . . . 5.2.3 The spaces Vm for shells . . . . . . . . . . . . . . . . . . . . 5.2.4 Direct problems for nonfastened plates and shells . . . . . . 5.2.5 Solution of optimization problems . . . . . . . . . . . . . . 5.3 Optimization problems for plates (control by the function of the thickness) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1 Optimization under restrictions on strength . . . . . . . . . 5.3.2 Stability optimization problem . . . . . . . . . . . . . . . . 5.3.3 Optimization of frequencies of free oscillations . . . . . . . . 5.3.4 Combined optimization problem and optimization for a class of loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Optimization problems for shells (control by the functions of the midsurface and of the thickness) . . . . . . . . . . . . . . . . . . .

291

291 291 293

295 295 296 298 301 302 305 305 310 315 316 317

xii

CONTENTS 5.4.1 Problem of optimization of a shell of revolution with respect to the strength . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.2 Optimization according to the stability of a cylindrical shell subject to a hydrostatic compressive load . . . . . . . . . . Control by the shape of a hole and by the function of thickness for a shallow shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.1 Problem of optimization according to the strength . . . . . 5.5.2 Approximate solution of the optimization and direct problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.3 Problem of optimization of eigenvalues . . . . . . . . . . . . 5.5.4 Approximate solution of the eigenvalue problem . . . . . . Control by the load for plates and shells . . . . . . . . . . . . . . . 5.6.1 General problem of the control by the load . . . . . . . . . 5.6.2 Optimization problems for plates . . . . . . . . . . . . . . . Optimization of structures of composite materials . . . . . . . . . . 5.7.1 Concept of a composite material . . . . . . . . . . . . . . . 5.7.2 Homogenization (averaging) of a periodical structure based on the G-convergence . . . . . . . . . . . . . . . . . . . . . 5.7.3 E ective elasticity characteristics of granule and ber reinforced composites . . . . . . . . . . . . . . . . . . . . . . . . 5.7.4 Optimization of the e ective elasticity constants of a composite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.7.5 Optimization of a granule reinforced composite . . . . . . . 5.7.6 Optimization of composite laminate shells . . . . . . . . . . 5.7.7 Optimization of the composite structure . . . . . . . . . . . Optimization of laminate composite covers according to mechanical and radio engineering characteristics . . . . . . . . . . . . . . . . . 5.8.1 Propagation of electromagnetic waves through a laminated medium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.8.2 Optimization problems . . . . . . . . . . . . . . . . . . . . . Shape optimization of two-dimensional elastic body . . . . . . . . . 5.9.1 Sets of controls and domains in the optimization problem . 5.9.2 Problems of elasticity in domains . . . . . . . . . . . . . . . 5.9.3 The optimization problem . . . . . . . . . . . . . . . . . . . Optimization of the internal boundary of a two-dimensional elastic body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Optimization problems on manifolds and shape optimization of elastic solids . . . . . . . . . . . . . . . . . . 5.11.1 Optimization problem for an elastic solid . . . . . . . . . . 5.11.2 Spaces and operators on R=2 Z, auxiliary statements . . . 5.11.3 Optimization problem on R=2 Z . . . . . . . . . . . . . . . Optimization of the residual stresses in an elastoplastic body . . . 5.12.1 Force and thermal loading of a nonlinear elastoplastic body 5.12.2 Residual stresses and deformations . . . . . . . . . . . . . . 317 320 323 323 326 328 329 330 330 331 337 337 338 347 353 359 362 371 378 378 385 387 388 389 390 393 396 397 403 410 414 415 426

5.5

5.6 5.7

5.8

5.9 5.10 5.11

5.12

CONTENTS

xiii

5.12.3 Temperature pattern in a medium . . . . . . . . . . . . . . 430 5.12.4 Optimization problem . . . . . . . . . . . . . . . . . . . . . 431

6 Optimization problems for steady ows of viscous and nonlinear viscous uids 435
6.1 Problem on steady ow of a nonlinear viscous uid . . . . . . . . . 6.1.1 Basic equations and assumptions . . . . . . . . . . . . . . . 6.1.2 Formulation of the problem . . . . . . . . . . . . . . . . . . 6.1.3 Existence theorem . . . . . . . . . . . . . . . . . . . . . . . 6.2 Theorem on continuity . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Continuity with respect to the shape of the domain . . . . . . . . . 6.3.1 Formulation of the problem . . . . . . . . . . . . . . . . . . e e 6.3.2 Lemmas on operators Lq and Bq . . . . . . . . . . . . . . . 6.3.3 Theorem on continuity . . . . . . . . . . . . . . . . . . . . . 6.4 Control of uid ows by perforated walls and computation of the function of ltration . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.1 The problem on ow in a circular cylinder and the function of ltration . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4.2 The passage factor for the power model . . . . . . . . . . . 6.4.3 Control of the surface forces at the inlet by the perforated wall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5 The ow in a canal with a perforated wall placed inside . . . . . . 6.5.1 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.2 Generalized solution of the problem . . . . . . . . . . . . . 6.6 Optimization by the functions of surface forces and ltration . . . 6.6.1 Formulation of the problem and the existence theorem . . . 6.6.2 On the di erentiability of the function T ! (v(T ) p(T )) . . 6.6.3 Di erentiability of the functionals i and necessary optimality conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7 Problems on the optimal shape of a canal . . . . . . . . . . . . . . 6.7.1 Set of controls and di eomorphisms . . . . . . . . . . . . . 6.7.2 Optimization problems . . . . . . . . . . . . . . . . . . . . . 6.8 A problem on the optimal shape of a hydrofoil . . . . . . . . . . . 6.8.1 State equation for a moving hydrofoil . . . . . . . . . . . . 6.8.2 Fixed-domain problem and Frechet di erentiability of the functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.8.3 Optimization problem . . . . . . . . . . . . . . . . . . . . . 6.9 Direct and optimization problems with consideration for the inertia forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9.1 Setting and solution of the direct problem . . . . . . . . . . 6.9.2 Approximation of the problem (6.9.10){(6.9.12) . . . . . . . 6.9.3 Some remarks on models, optimization problems, and existence results . . . . . . . . . . . . . . . . . . . . . . . . . .

435 435 438 443 447 451 451 453 456

459 459 462 463 465 465 466 468 468 470 474 476 477 479 483 483 490 498 500 501 505 506

xiv

CONTENTS

Bibliography

509

Preface
This book is intended to be both a thorough introduction to the contemporary research in the optimization theory for elliptic systems with its numerous applications and a textbook at the undergraduate and graduate level for courses in pure or applied mathematics or in continuum mechanics. Various processes of modern technology and production are described by elliptic partial di erential equations. Optimization of these processes reduces to optimization problems for elliptic systems. The numerical solution of such problems is associated with the solution of the following questions. 1. The setting of the optimization problem ensuring the existence of a solution on a set of admissible controls, which is a subset of some in nite-dimensional vector space. 2. Reduction of the in nite-dimensional optimization problem to a sequence of nite-dimensional problems such that the solutions of the nite-dimensional problems converge, in a sense, to the solution of the in nite-dimensional problem. 3. Numerical solution of the nite-dimensional problems. The book is devoted to these questions. Attention is focused on the setting of a problem, on the proving of an existence theorem, and on methods of approximate solution of optimization problems, the existence theorems being the base for the construction of approximate solutions. General systems of elliptic equations are considered. For such systems, we study optimization problems in which the coefcients of equations, the shape of domain, and the right hand sides of equations are considered to be controls. General results are applied to various optimization problems of mechanics of deformable bodies, plates, shells, composite materials, and structures made of them, as well as to optimization problems of mechanics of viscous and nonlinear viscous uids. In order to read the book it is necessary to know some basic topics of topology, function theory, and functional analysis. Nevertheless, all the necessary mathematical topics are set forth in Chapter 1. xv

xvi

Preface

I hope the book will be useful not only for students and scientists in mathematics and mechanics, but also for researching engineers and various applicationminded readers. It will be possible for them to see correct formulations of direct and optimization problems, methods of approximation, and possibly they will be convinced that modern mathematical analysis is an indispensable tool for the setting, genuine understanding, and solution of the direct and optimization problems arising in applications. Mathematically oriented readers will see many open mathematical problems issued from technology and production. Chapters 1 and 4 of this book were the basis of my lectures to undergraduate students of Kiev University. I must acknowledge the contributions of the many people who aided, either directly or indirectly, in the conception and writing of the manuscript. The seminars of Professors Yu. M. Berezansky, S. D. Eidelman, and M. L. Gorbachuk attached to the Institute of Mathematics of National Academy of Sciences of Ukraine have exerted an essential in uence on the creation of this book. The attention and encouragement of the leaders and participants of the seminars stimulated my activity. Some parts of the book entered into my lectures to graduate students of the University of Jyvaskyla. I very sincerely thank Professor P. Neittaanmaki for the invitation and hospitality. I had the pleasure of delivering lectures on topics of the book in the Voronezh Winter Mathematical Schools, whose organizer, life, and soul was Professor S. G. Krein, and in the Crimean Fall Mathematical Schools, organized by Professor N. D. Kopachevsky. Many famous mathematicians participated in these schools, and I had a good chance to discuss various problems with them. Discussions with Professors S. G. Krein and P. E. Sobolevsky were especially helpful to me. Some of the problems considered in the book have appeared as a result of discussions with Professors N. V. Banichuk and K. A. Lurie. Professors N. S. Bakhvalov, A. V. Fursikov, G. M. Kobelkov, V. B. Kolmanovsky, R. Mennicken, Z. Mroz, J. Necas, E. Schnack, I. V. Skrypnik, D. Tiba, V. M. Tikhomirov o ered me opportunities of delivering lectures, which gave me a clearer view on the subject, and I could test various forms of the presentation of the material of the book. I am gratefully indebted to Professors Yu. Deynekin, Ya. Rojtberg, A. Seyranian, Z. Sheftel for many helpful discussions. During many years I had the fortune of working in the Institute of Mechanics of National Academy of Sciences of Ukraine with my colleagues Professors I. Babich, A. Panteleev, Yu. Rubezhanky, N. Semeniuk, Doctors M. Belonosov, N. Ivanova, T. Zarutskaia. Each of them facilitated the creation of the book. Since May 1997, I live in Germany and here I had plenty of useful discussions with Professors K.-H. Ho mann, R. Hoppe, B. Kroplin, E. Meister, and W. Wendland. Moreover, these professors, especially Prof. B. Kroplin, have made many good deeds for me. I am very thankful to the editor of the series Operator Theory : Advances and

Preface

xvii

Applications Prof. I. Gohberg for the amiable consent to publish the manuscript of the book. I am also indebted to Dr. A. Butyrin and my son, Dr. E. Lytvynov, for the translation into English of the most part of the manuscript written in Russian, and for signi cant improvements of my English in the other part, as well as for working up the manuscript into print. I very sincerely thank Dr. N. Botkin, Dr. G. Shchepan'uk, and Eng. T. Maier for the help in the preparation of the book for publication. Stuttgart, April 1999 William G. Litvinov

xviii

Preface

Introduction
\This, O my Best Beloved, is a story|a new and a wonderful story|a story quite di erent from the other stories|a story about the Most Wise Sovereign Suleiman-bin-Daoud-Solomon the Son of David. \There are three hundred and fty- ve stories about Suleiman-bin-Daoud but this is not one of them." | R. Kipling \The Butter y that Stamped" Optimization problems for elliptic equations are met with in various applications. A great number of such problems arise in continuum mechanics, in the optimization of structures, processes, and so forth. The book is devoted to the optimization problems for elliptic equations and systems of equations. In the majority of cases, we apply the generalized (weak) solution of a boundary value problem, which is de ned as an element u such that

u2V

a(u v) = (f v)

for all v 2 V:

(1)

Here, V is a Hilbert space, a is a bilinear form generated by an elliptic equation or by a system of elliptic equations, and f is the right hand side, which is considered to be an element of the dual V of the space V . In problems of mechanics of deformable bodies, the bilinear form a is generated by the strain energy, and the right hand side f by the applied load. From the physical point of view, the equation (1) is the principle of virtual work. Analogously, a generalized solution of an eigenvalue problem is de ned as a set of pairs (ui i ) V R such that

a(ui v) = i b(ui v) for all v 2 V: (2) Here, b is a bilinear form on V V , R is the set of real numbers, b is generated by
the kinetic energy of the deformable body, the equation (2) is obtained from the xix

xx

Introduction

principle of virtual work in which the inertial forces are considered to be applied loads, i being connected with the frequencies of free vibrations. The bilinear forms a and b depend on a control, e.g., in the case of a shell, the functions of thickness and curvatures of the midsurface may be considered as a control, the bilinear forms a, b depending on these functions. In order to cover a wide range of optimization problems, we consider the Q following abstract bilinear form a given on the space W = i=1 W2li ( ), where W2li ( ) is the Sobolev space, li 1, V W , and

a(u v) =

Z X k

Here, Pi 's are linear continuous mappings of W into L2 ( ), and the variable coefcients aij depend on a control. In some cases, the operators Pi may also depend on a control. In problems of mechanics of deformable bodies, Pi u's are the components of the strains generated by a function of displacement u. The goal and the restriction functionals, the sets of controls and admissible controls are introduced. The goal and restriction functionals are de ned on the set of the pairs (h uh), where h is a control and uh a solution of the problem (1) or (2) for the control h. Usually, the set of controls is supposed to be a convex set in a Sobolev space or in the product of Sobolev spaces, and these spaces are chosen so that the goal and restriction functionals are continuous with respect to the topology generated by the weak one of the corresponding Sobolev space (or of the product of Sobolev spaces). The set of admissible controls is given by putting restrictions on the values of the restriction functionals and on some characteristics of h and uh . The optimization problem consists in nding an element that minimizes or maximizes the goal functional on the set of admissible controls. For problems under consideration, we prove existence theorems and study some methods of the construction of approximate solutions. We also consider the domain shape optimization problems in which a control de nes the shape of domain. Here, we apply smooth di eomorphisms of domains de ned by the set of controls onto a xed domain, and after the replacement of variables corresponding to the di eomorphisms, the domain shape optimization problem reduces to a problem of control by coe cients in a xed domain. We also study optimization problems in which the right hand sides of equations are considered to be controls. Here, existence theorems are proved, necessary optimality conditions are established, and methods of approximation are studied. General results are applied to exploring and solving various optimization problems of mechanics of deformable bodies and mechanics of viscous and nonlinear viscous uids, including various problems of technology and production. A typical peculiarity of almost all problems considered in the book is that the set of admissible controls is nonconvex. Moreover, the goal functional is also nonconvex in some problems. These circumstances create considerable di culties for the numerical solution of optimization problems. Besides, the direct problems of

i j =1

aij (Pi u)(Pj v) dx:

Introduction

xxi

deformable bodies and viscous and nonlinear viscous uids reduce to complicated systems of elliptic equations. This is why only few numerical solutions of nonconvex optimization problems are presented in the book, and the solutions found are locally optimal or the best of some locally optimal solutions. It should be noted that, when the coe cients of equations or the shape of domain are considered to be a control, we apply, as a rule, controls of Sobolev spaces, which possess some smoothness. In many cases, smooth controls comply with the physical and engineering points of view. For example, the models of plates and shells are valid if the derivatives of the function of thickness are not great and the functions of curvatures of a shell are smooth. Moreover, in case of a nonsmooth function of thickness, the concentration of stresses, leading to the failure of a structure, arises, and this concentration is not taken into account in the models of plates and shells. Similar situations take place in optimization of various processes of technology and production. Discontinuous controls were calculated for the control of ight of rockets. However, the engineers rejected these, as such controls might lead to dangerous e ects, which were not taken into account in the mathematical models. \The control as well as the management have to be gentle," as the saying goes. We note also that, in the majority of the general optimization problems studied in the book, the boundary value problems are considered in bounded domains in R2 . Nevertheless, almost all the results obtained may be extended to bounded domains in Rn with n > 2. The book consists of six chapters. In Chapter 1, we set forth some topics from functional analysis needed to read the book, and establish some results which are used in the following chapters for the investigation and solution of the optimization problems. We introduce the concepts of topological, metric, normed, and Hilbert spaces, linear and k-linear continuous mappings. Spectral problems associated with bilinear continuous forms in a Hilbert space are considered. The Sobolev spaces are de ned, and the imbedding theorem, the theorem on the coerciveness of a system of operators, the implicit function theorem, and others are adduced. The problem of the di erentiability of the eigenvalues of (2) with respect to the control that determines the forms a and b is studied. Chapter 2 is devoted to the optimization by coe cients of elliptic systems. We consider optimization problems in which the state of a system is de ned by one function or by a set of functions. The latter is an abstract analog of the problem of optimization of a structure subject to a class of loads. Various problems of eigenvalue and domain shape optimization are studied. In Chapter 3, we study optimization problems in which the right hand sides of equations are controls. Existence theorems and the necessary optimality conditions are established, and the convergence of approximate solutions is studied. We also consider the minimax optimization problems and problems of optimization of systems described by variational inequalities. In Chapter 4, we study direct problems for various models of plates and shells, problems of stress-strain state, of stability, and of free vibrations. The bilinear

xxii

Introduction

forms associated with the potential (strain) and kinetic energies of plates and shells are de ned and studied. Existence results for the problems under consideration are established. The reader interested in the models and mathematical theory of plates and shells can read this chapter after looking through Sections 1.1{1.9. Chapter 5 is devoted to the optimization of structures. We consider various settings of optimization problems and numerical methods for solving direct and optimization problems. Optimization problems for plates and shells under restrictions on strength, sti ness, stability, and frequencies of free vibrations are studied. We also consider various problems of optimization of composite materials and structures made of them. In Chapter 6, we formulate and study various optimization problems for steady ows of viscous and nonlinear viscous uids. In particular, we consider the optimization by body and surface forces, by the distribution of velocities on the inlet, by perforated walls, and by the shape of domain. We also study some engineering problems, such as problems of the optimal shape of a canal and of the optimal shape of the hydrofoil and the problem of optimization of the header of a paper machine.

Chapter 1

Basic de nitions and auxiliary statements

\All this recitative by the chorus is only to bring us to the point where you may be told why Dry Valley Johnson shook up the insoluble sulphur in the bottle" | O. Henry \The Indian Summer of Dry Valley Johnson"

1.1 Sets, functions, real numbers

1.1.1 Notations and de nitions
We suppose the reader to be familiar with notions connected with sets and functions nevertheless, we state the terminology and introduce notations that will be used throughout the book. Let X be a set the writing x 2 X means that x is an element of the set X if the writing x 2 X follows some equality (or inequality) and no additional conditions are imposed on x, then it should be read that the equality (inequality) holds for an arbitrary x from X . The sign 8 means \for all." In the case when it is clear which set X we bear in mind, we simply write 8x instead of 8x 2 X . The writing x 62 X means that x does not belong to X . The writing fx j P (x)g denotes the set of all the x satisfying the condition (or conditions) P (x). If A is a subset of a set X , which is denoted by A X , then X n A is the complement of the subset 1

Chapter 1. Basic de nitions and auxiliary statements

A to the whole set X , i.e., X n A = x j x 2 X x 62 A :

The product E F of two sets E and F is the set of all ordered pairs (x y) such that x is an element of the set E and y an element of the set F . The word \ordered" means that, if E and F coincide, then the pairs (x y) and (y x) with x 6= y are considered as di erent ones. In the same way, one de nes the product of several sets, or of an arbitrary collection of sets. Let sets E and F be given. A map (mapping) of E into F , or a function de ned on E and taking values in F , or an operator acting from E into F , is a correspondence f that assigns to every element x from E some element from F , which is denoted by f (x) (or fx) and is called the image of x under the mapping f . In particular, if F is the set of real numbers, then f is called a functional. The notation f : E ! F means that f is a mapping of E into F . A mapping f is said to be injective, or an injection, if the images of two di erent elements of E under the mapping f are di erent elements of F . A mapping f is called surjective, or a surjection, or a mapping \onto" if every element of F is the image under the mapping f of some element of E . A mapping f : E ! F is called bijective, or a bijection, or a one-to-one mapping, if every element of F is the image of a unique element of E . A mapping is one-to-one if and only if it is both injective and surjective. Let f : E ! F be a one-to-one mapping and let y belong to F . Denote by f ;1 (y) the unique element x of E such that f (x) = y. So, the mapping f ;1 from F into E is well-de ned. It is also a one-to-one mapping. This mapping is called the inverse of f . The set of real numbers is denoted by R, the set of integers by Z, and the set of natural numbers by N . The set of real x satisfying a < x < b (a, b are real numbers, or it may be that a = ;1, b = 1) is called an open interval and is denoted by (a b). The set of those x which satisfy the inequalities a x b is called a segment, or a closed interval, and is denoted by a b]. We will use also the following notations: a b) = fx j a x < bg, (a b] = fx j a < x bg, R+ = 0 1). Sets a b), (a b], where a b < 1, are called half-intervals. Let A be a subset of R. The set A is called bounded above (below) if there exists a number M (m) that is not less (greater) than all the numbers x from A, M called a majorant of the set A, and m a minorant. A set A is called bounded if it is both bounded above and below. A set in R is said to have maximum if there exists a majorant that belongs to this set. A set The union of sets A, B is denoted by A B , and the intersection of A, B by

A \ B.

1.1.2 Real numbers

1.1. Sets, functions, real numbers

a and b are two maximums of the same set, then we have both a b and b a, so a = b. The maximum of a set A (if it exists) is denoted by max x. x2A

need not have maximum, but if it does have, then this maximum is unique. For if

The set that is the union of the real numbers R and the set consisting of the two elements ;1, +1 is called the completion of the set of real numbers R, and is denoted by R . This set has minimum ;1 and maximum +1. Every set in R is bounded, and Theorem 1.1.1 holds true in R without the assumption that A is majorized (or minorized), since this is always the case. Let us introduce the notions of upper and lower limits. Let fxn g be a sequence of elements from R. Denote by lim ptfxn g the collection of the limit points of the sequence fxn g. Then, the upper limit L of the sequence fxn g is de ned by L = sup x j x 2 lim ptfxn g : It is denoted by L = lim supfxn g. Analogously, we can de ne a lower limit l: l = lim inf fxn g = inf x j x 2 lim ptfxn g

An analogous de nition is stated for the minimum, which is denoted by min. A set A in R is said to have supremum if the set of its majorants has minimum, and this minimum is called the supremum of the set A. So, the supremum is the least of the majorants. A set need not have supremum, but if it has, then this supremum is unique. The supremum of a set A (if it exists) is denoted by sup x. x2A If the supremum belongs to A, then it is the maximum, and vise versa. An analogous de nition is stated for the in mum, which is denoted by inf. If the in mum belongs to A, then it is the minimum. The following theorem holds, see, e.g., Bourbaki (1960). Theorem 1.1.1 An arbitrary, nonempty, majorized set in R has supremum, and an arbitrary, nonempty, minorized set has in mum. The supremum b of a majorized set A is characterized by the following statements : i) x b for any x 2 A ii) for an arbitrary b1 such that b1 < b, there exists at least one number x 2 A such that b1 < x b. The in mum c of a minorized set A is characterized by the following statements : i) x c for any x 2 A ii) for an arbitrary c1 such that c1 > c, there exists at least one number x 2 A such that c1 > x c: For example, sup x = 0 inf x = 0 sup x = max x = 0 x2(0 1) x2(;1 0] x2(;1 0) x2(;1 0] inf1) x = x2 0 1) x = 0: min x2 0

Chapter 1. Basic de nitions and auxiliary statements

The upper limit is characterized by the following property. For arbitrary L1 and L2 such that L1 < L < L2, all xn but a nite number are less than or equal to L2 , and there is an in nite set of n such that xn L1 . We stress that every bounded sequence in R has supremum and in mum. A sequence in R converges if and only if the upper and lower limits of it are equal.

1.2.1 General notions

Topological spaces

1.2 Topological, metric, and normed spaces

A topological space X is a set X in which there has been chosen a collection of subsets, called open sets, that have the following properties: a) the space X itself and the empty set ? are open b) the intersection of a nite number of open sets is open c) the union of either a nite or in nite number of open sets is open d) for arbitrary di erent points a and b of the space X , there exist two open disjoint sets from X that contain a and b, respectively. Notice that open sets are usually supposed to satisfy only the conditions a){ c). If, additionally, the condition d) is satis ed, then the topological space is called a Hausdor space. The collection is called a topology in X . One set may be supplied with di erent topologies, so that we will indicate the topology with respect to which some property holds. The collection of all topologies on the set X is naturally ordered: 1 2 if 1 2 (in the sense of the usual inclusion). If 1 2 , then the topology 1 is said to be weaker than the topology 2 . (The word \weaker" is used because more sequences converge in the topology 1 than in the topology 2 , so that the 1 -convergence is a weaker concept than the 2 -convergence.) Let A be a set in a topological space X . The set A is called closed if its complement is open. The closure A of the set A is the intersection of all the closed sets containing A. The interior of the set A is the union of all the open sets contained in A. The interior of the set A is denoted by A A is the biggest open set contained in A A may be empty. The set A n A is called the boundary of A. A neighbourhood of a point a in the set X is an arbitrary subset of X containing an open set the point a belongs to. A collection S of subsets of a topological space X is called a base of neighborhoods of a point a if each N 2 S is a neighborhood of a and, for an arbitrary neighborhood M of a, there exists N 2 S such that N M . A set M in a topological space X is called dense in X if its closure coincides with X , i.e., M = X . A topological space is called separable if it has a countable

1.2. Topological, metric, and normed spaces

dense set. A sequence fxn g of points of a topological space X converges in X to some x 2 X if, for an arbitrary neighborhood U of x, there exits a natural number N (U ) such that all xn , with n > N (U ), belong to U . If a sequence fxn g converges to x, then x is called its limit, and one writes xn ! x in X .

Induced topology. Product of topologies

Let E be a subset of a topological space X . On the set E , one can introduce a topology by taking as open sets the intersections of the open sets in X with E . Then, E is called the topological subspace of X , or the topological space with the topology induced (generated) by the topology of X . Let X1 , X2 be two topological spaces. On the product of these spaces, one can introduce a topology in the following way. A set E in X1 X2 is called open if, for each point (x1 x2 ) 2 E , there are open sets A1 , A2 in X1 , X2 , respectively, such that x1 2 A1 , x2 2 A2 , and A1 A2 E . It can be veri ed that this de nition is correct, i.e., the open sets de ned satisfy all the axioms of a topological space. As easily seen, the product of two arbitrary open sets is again an open set, though there are also other open sets. Such a topology on X1 X2 is called the product of the topologies of X1 and X2 , or the topological product of X1 and X2 . In the same way one de nes the topological product of more than two spaces. For example, if X1 = X2 = = Xn = R, then their product is just the space Rn equipped with the natural metric.

1.2.2 Metric spaces

A set E is called a metric space if the distance between its elements is de ned, i.e., there is a mapping of E E into the half-line R+ that places in correspondense to an arbitrary pair (x y) 2 E E a number d(x y), which is called the distance between x and y. The distance must posses the following three properties: 1. the symmetry: d(x y) = d(y x) 2. the positivity: d(x y) > 0 if x 6= y, and d(x x) = 0 3. the triangle inequality: d(x z ) d(x y) + d(y z ). In particular, the real line R with the distance d(x y) = jx ; yj is a metric space. This metric is called the natural metric in R. Everywhere below, the metric in R is supposed to be natural. Let E be a metric space. Given an arbitrary point a from E , the set

x j x 2 E d(a x) < r

Chapter 1. Basic de nitions and auxiliary statements

is called the open ball centered at a with radius r, and the set

x j x 2 E d(a x) r
the closed ball centered at a with radius r, r being nite and nonzero. A set A in the metric space E is called open if, together with its every point, it contains a ball (open or closed) centered at this point. Clearly, open sets in the space E possess the properties a), b), c), and d) stated above. Therefore, a metric space is a speci c case of a topological space, and its topology is determined by the metric. A topological space is said to be metrizable if there exists a metric generating its topology. A sequence of points fxn g of a metric space E converges to a point x 2 E in the sense of the topology of this space if and only if nlim d(xn x) = 0 Notice !1 that, because of the triangle inequality, d(xm xn ) ! 0 as m n ! 1. A sequence possessing this property is called a Cauchy sequence (or a fundamental sequence). A metric space E is called complete if every fundamental sequence in it converges to an element of E .

1.2.3 Normed vector spaces

Basic de nitions
A set E is called a vector space over the eld of real numbers R if given are two operations on it, named multiplication by a scalar and addition. These operations are supposed to possess the following algebraic properties:

(i) to every pair of vectors (elements) x y 2 E , there corresponds a vector x + y 2

E , and the equalities x+y =y+x x + (y + z ) = (x + y) + z
(1.2.1) hold E contains a unique vector 0 (the zero vector) such that x + 0 = x for any x from E for every x 2 E , there exists a unique vector ;x such that x + (;x) = 0

(ii) to every pair (t x), where t 2 R and x 2 E , there corresponds a vector tx 2 E ,

and the equalities 1x = x s(tx) = (st)x s 2 R s(x + y) = sx + sy (s + t)x = sx + tx: hold.

1.2. Topological, metric, and normed spaces

A segment with ends x and y in the vector space E is the set of the elements of the form tx + (1 ; t)y, 0 t 1. It is denoted by x y]. A set V E is called convex if, for every pair (x y) of its elements, the segment x y] belongs to V . A cone K E is de ned to be a set containing, together with its arbitrary element x, the elements of the form tx, where t > 0. Note that 0|the zero element of E |may either belong or do not belong to K . Provided 0 2 K , K is called a cone with vertex at 0. A cone that is a convex set is called a convex cone. Note that a cone need not be \sharpened." For example, an arbitrary subspace of E is a cone. Here are two important examples of convex cones in the case E = Rn . The nonnegative orthant:

x = (x1 : : : xn ) j x1 0 : : : xn 0 :
The positive orthant:

x = (x1 : : : xn ) j x1 > 0 : : : xn > 0 :

Notice that the rst example is a cone with vertex at 0. Let f be a mapping of a convex set V E into R. This mapping is called convex if, for arbitrary x y 2 V and an arbitrary 2 0 1], the following inequality holds: f ( x + (1 ; )y) f (x) + (1 ; )f (y): A norm in the vector space E is de ned to be a function x ! kxk that assigns a number kxk to a vector x and possesses the following properties: a) kxk > 0 if x 6= 0, and k0k = 0 b) kx + yk kxk + kyk for any x y 2 E c) ktxk = jtjkxk for any x 2 E and t 2 R, In a normed vector space, one can de ne a distance between elements satisfying the properties 1), 2), 3) above by the formula d(x y) = kx ; yk. Therefore, an arbitrary normed vector space is a metric space. In a normed vector space, one can introduce the notion of a completeness (see subsec. 1.2.2). A complete normed vector space is called a Banach space. Examples of Banach spaces are `1 and `1 0 , which will be used in the sequel. `1 is the space of bounded number sequences. The norm of an element = ( 1 2 : : : n : : : ) 2 `1 is de ned by

k kl1 = sup j i j
i

`1 0 is the subspace of `1 consisting of the sequences converging to zero. The norm in `1 0 is de ned just as in `1 .

Chapter 1. Basic de nitions and auxiliary statements

Two norms k k1 and k k2 de ned in a vector space E are called equivalent if there exist constants m1 and m2 such that

kxk1 m1 kxk2 kxk2 m2 kxk1 x 2 E: The metrics d1 (x y) = kx ; yk1 and d2 (x y) = kx ; yk2 corresponding to equivalent norms de ne in E identical systems of open sets, i.e., they de ne identical topologies.

Theorem 1.2.1 In a nite-dimensional vector space, arbitrary two norms are equivalent. There exists one system of open sets for any norm introduced in this space.
For the proof, see, e.g., Schwartz (1967). For example, consider the nite-dimensional space Rn . A norm on Rn can be de ned in di erent ways. For example, if x = (x1 x2 : : : xn ), then

kxk =

n 2 X 2 1 xi : i=1

This norm is called Euclidean. It P possible also to introduce a norm by the is formulas kxk = maxi jxi j or kxk = n=1 jxi j. All these norms are equivalent. i Let X and Y be vector normed space, and let A be a linear mapping of X into Y , i.e.,

A( x + y) = Ax + Ay x y2X 2 R: The mapping A is called bounded if there exists a constant c such that kAxkY ckxkX x 2 X: Denote by L(X Y ) the set of all bounded linear mappings of X into Y , in the case when X = Y , we use the notation L(X Y ) = L(X ). On the set L(X Y ), one can
naturally de ne the addition of elements and the multiplication by a scalar. So, L(X Y ) is a vector space. In this space, we de ne a norm as follows:

Ax kAk = sup kkxkkY = sup kAxkY

x6=0 X
kxkX =1

for an arbitrary element A 2 L(X Y ). In the case Y = R, the space L(X R) is denoted by X and is called the dual of X . Elements of X are called bounded linear functionals in X , or bounded linear forms in X . The value of a linear functional f from X at an element x 2 X is usually denoted by (f x) the function ( ), de ned on X X , being called the scalar product of X and X .

1.2. Topological, metric, and normed spaces

Theorem 1.2.2 If E , F are normed vector spaces and F is a Banach space, then the space L(E F ) is also a Banach space. In particular, the dual E of the space
E is a Banach space.
For the proof, see, e.g., Schwartz (1967). Let X be a Banach space, and let X be its dual, which is also a Banach space. The space X , in turn, has its dual, which is denoted by X and is called the second dual of X . For an arbitrary element x 2 X , there exists a unique element x 2 X possessing the property (x x ) = (x x)

x 2X :

Moreover, kxkX = kx kX . So, the space X may be considered as a subspace of the Banach space X . A Banach space X is called re exive if X = X .

Weak topology and weak convergence in a Banach space

In addition to the topology generated by the norm and called the strong topology, one introduces in a Banach space a weak topology. The weak topology in a Banach space is the weakest topology on X with respect to which every functional f 2 X is continuous. For the weak topology, a base of neighborhoods of zero is given by the following sets

N (f1 f2 : : : fn ") = x j x 2 X jfi (x)j < " i = 1 2 : : : n :

Here f1 f2 : : : fn is a nite collection of arbitrary elements of X , " is a positive number. The convergence with respect to the strong topology in a Banach space X is called strong convergence, or the convergence in norm. Everywhere below, if we deal with a a normed space and do not indicate which topology is under consideration, then the strong topology is understood. The convergence with respect to the weak topology is called weak convergence a sequence fxn g of elements of X converges weakly to x0 2 X if and only if, for an arbitrary f 2 X , the number sequence ff (xn )g converges to f (x0 ). The following theorem holds (see, e.g., Kantorovich and Akilov (1977)).

Theorem 1.2.3 In a normed vector space X , the weak topology coincides with the strong topology if and only if the space X is nite-dimensional.
We stress that, in an in nite-dimensional Banach space, the weak topology is not generated by any metric. Because of convexity of a norm, we have the following theorem, see Vainberg (1972), Gajewski et al. (1974).

10
then

Chapter 1. Basic de nitions and auxiliary statements

Theorem 1.2.4 If a sequence fxn g in a Banach space X weakly converges to x,

kxk lim inf kxn k: n!1
A sequence fxn g is called weakly fundamental if, for any functional f 2 X , there exists a nite limit limn!1 f (xn ). The space X is called sequentially weakly complete if every weakly fundamental sequence weakly converges to an element of X.
plete.

Theorem 1.2.5 An arbitrary re exive Banach space is sequentially weakly comFor the proof, see, e.g., Yosida (1971). In a normed vector space X , one can introduce the notions of a weakly closed set and a sequentially weakly closed set. A set A X is called weakly closed if it is closed in the weak topology of X . A set A is called sequentially weakly closed if the conditions fxn g A and xn ! x weakly in X yield x 2 A. A weakly closed set is sequentially weakly closed. However, in general, the inverse statement is not valid. Note that an arbitrary closed (in the sense of the strong topology) convex set in a Banach space is sequentially weakly closed.

1.3 Continuous functions and compact spaces

1.3.1 Continuous and semicontinuous mappings
The continuity
Let E , F be topological spaces, let f be a mapping of E into F , and let A be a subset of E . By f (A) we denote the subset of F consisting of all the elements f (x), x 2 A f (A) is called the image of the set A under the mapping f . Let B be a subset of F . By f ;1 (B ) we denote the set in E consisting of all elements x such that f (x) 2 B . The set f ;1 (B ) is called the prototype (or the inverse image) of the set B under the mapping f . A mapping f of a topological space E into a topological space F is continuous at a point a 2 E if, for any neighbourhood U of the point f (a), there exists a neighborhood V of the point a such that f (V ) U . A mapping of E into F is called continuous if it is continuous at every point a of the space E .

Theorem 1.3.1 A mapping f of a topological space E into a topological space F is continuous if and only if the prototype of an arbitrary open (closed ) set in F is an open (closed ) set in E .
For the proof, see, e.g., Schwartz (1967), Kolmogorov and Fomin (1975).

1.3. Continuous functions and compact spaces

11

In particular, a linear mapping A of a normed vector space X into a normed vector space Y is continuous if and only if it is bounded. Let F G be normed vector spaces, and let A be a linear continuous mapping of F into G, i.e., A 2 L(F G). This mapping is said to be invertible if it is a bijection and the inverse bijection A;1 is also continuous, i.e. A;1 2 L(G F ). Then, the mapping A is called an isomorphism. A homeomorphism of a topological space E onto a topological space F is de ned to be a bijection of E onto F such that both it and its inverse are continuous mappings.

Theorem 1.3.2 Let F , G be Banach spaces and let U (U ;1 ) be the set of the invertible elements of the space L(F G) (L(G F )). Then, the bijection f ! f ;1 of the set U onto the set U ;1 is a homeomorphism.
For the proof, see, e.g., Schwartz (1967). A mapping f of a topological space E into a topological space F is called sequentially continuous at a point a 2 E if the image of an arbitrary sequence of points from E converging to a is a sequence of points from F converging to f (a). A mapping of E into F is called sequentially continuous if it is sequentially continuous at every point of the space E . If topological spaces E and F are metrizable, then the notion of a continuous mapping is equivalent to the notion of a sequentially continuous mapping, that is, every continuous mapping of E into F is sequentially continuously, and vise versa. If E and F are topological nonmetrizable spaces, or so is at least one of them, then every continuous mapping of E into F is sequentially continuous, but the inverse is not true. Everywhere below, permitting ourselves some familiarity, instead of saying \f is a sequentially continuous mapping of E into F ," we will say \f is a continuous mapping of E into F ." Let f be a function de ned on a topological space E and taking values in R. This function is said to be upper semicontinuous at a point a 2 E if, for any given b1 > f (a), there exists in the space E a neighbourhood U 0 of the point a such that x 2 U 0 yields f (x) b1 . Analogously, f is lower semicontinuous at a point a if, for any given b2 < f (a), there exists a neighborhood U 00 in the space E of the point a such that x 2 U 00 yields f (x) b2 . A function f is continuous at a point if and only if it is both upper and lower semicontinuous at this point. A function f is called upper (lower) semicontinuous in E if it is upper (lower) semicontinuous at every point from E . If E is a metrizable topological space, then a function f : E ! R is upper semicontinuous if and only if for an arbitrary sequence fung convergent in E to u, the inequality lim sup f (un ) f (u) holds.
n!1
)

(1:3:1)

12

Chapter 1. Basic de nitions and auxiliary statements

Similarly, for a metrizable space E , a function f : E ! R is lower semicontinuous if and only if for an arbitrary sequence fung convergent in E to u, the (1:3:2) inequality lim inf f (un) f (u) holds. n!1 In the sequel, every function f : E ! R satisfying the condition (1.3.1) (or (1.3.2)) will be called upper (lower) semicontinuous even though the space E is nonmetrizable.

Bilinear and k-linear continuous mappings Let E , F , and G be vector spaces over R. A mapping x y ! u(x y) from the space
with respect to the other variable. If we x x 2 E , then u determines the partial linear mapping y ! u(x y) of the space F into G, which is denoted by u(x ). Analogously y 2 F being xed, the partial mapping x ! u(x y), denoted by u( y), is a linear mapping from E into G. In the case G = R, the mapping u is called a bilinear form on E F . The usual product of real numbers is a bilinear mapping of R R into R.

E F into the space G is called bilinear if, for any xed one variable, it is linear

Theorem 1.3.3 A bilinear mapping of a product E F of normed vector spaces

ku(x y)k M kxk kyk
x 2 E y 2 F:

into a normed vector space G is continuous if and only if there exists a constant M 0 such that

For the proof, see, e.g., Schwartz (1967). The set of the bilinear continuous mappings of E F into G is a vector space over R in which the norm is de ned by x kuk = sup kkux(k ky)kk = sup ku(x y)k: y
x6=0 y6=0
kxk

1 kyk 1

This space is denoted by L2 (E F G). Let E be a normed vector space, and let x y ! u(x y) be a bilinear form on E E , i.e., a bilinear mapping of E E into R. This form is called symmetric if

u(x y) = u(y x) x y 2 E: A form U is called coercive on E E if u(x x) ckxk2 x 2 E c = const > 0: Let now E1 E2 : : : Ek be vector spaces over R, and let E = E1 E2 Ek . A mapping u(x1 x2 : : : xk ) of E into G, where G is a vector space over R,

1.3. Continuous functions and compact spaces

13

is called k-linear if it is linear with respect to every xi 2 Ei . If Ei 's and G are normed vector spaces, then a k-linear mapping of E1 E2 Ek into G is continuous if and only if it is bounded, i.e., there exists a constant M 0 such that

ku(x1 x2 : : : xk )k M kx1 k kx2k kxk k

The set of k-linear continuous mappings of E1 vector space. The least constant M for which the latter inequality holds is the norm of the mapping u.

x = (x1 x2 : : : xk ) 2 E: E2 Ek into G is a normed

1.3.2 Compact spaces

Let E be a topological space. A collection of sets in E such that every point of E belongs to at least one of these sets is called a covering of E . A subcovering of a covering is a covering formed by some sets of the original covering. A covering is called nite if it consists of a nite number of sets in E . A covering is called open if all the sets of this covering are open in E . For example, the set of the intervals (n ; 1 n + 1), where n runs through the set of integers, forms an open covering of R. Note that it has no subcoverings. A topological space is called compact if, from every open covering of this space, one can choose at least one nite subcovering. A set A in a topological space E is called compact if, in the sense of the topology induced on A by the topology of E , A is a compact space. Notice that, in the latter case, the space E itself may be not compact. Every set in E whose closure is compact is called relatively compact.

Theorem 1.3.4 Let E be a topological space, and let F be a compact set in E .

Then, F is a closed set in E .

Theorem 1.3.5 The topological product of two compact spaces is compact.

A topological space E is called sequentially compact if, from an arbitrary sequence of elements of E , one can choose a subsequence converging to an element of E . We stress that, upon the Bolcano{Weierstrass theorem, for the metrizable topological spaces, the notions of compact and sequentially compact spaces are equivalent, that is, if E is a compact, metric space, it is also sequentially compact, and conversely a sequentially compact, metric space is a compact, metric space. If E is a nonmetrizable compact topological space, then it is sequentially compact, too. However, the inverse is not true.

Theorem 1.3.6 A set in a nite-dimensional normed vector space is compact if

and only if it is closed and bounded.

For the proofs of Theorems 1.3.4{1.3.6, see, e.g., Schwartz (1967). The following theorem is a very important property of a re exive space.

14

Chapter 1. Basic de nitions and auxiliary statements

Theorem 1.3.7 Every closed ball K in a re exive Banach space E is sequentially compact in the weak topology, that is, from every sequence of elements of K one can choose a subsequence weakly converging to an element of K .
For the proof, see Cea (1971).

1.3.3 Continuous functions on compact spaces

Theorem 1.3.8 Let f be a continuous mapping of a topological compact space E

For the proof, see, e.g., Schwartz (1967). The following theorem will be of great importance for us.

into a topological space F . Then, the image f (E ) of the space E under the mapping f is compact in F .

Theorem 1.3.9 (Weierstrass) A continuous mapping of a nonempty compact space into R reaches its maximum and minimum. Proof. Let F be a continuous mapping of a compact space E into R. Then, by Theorem 1.3.8, the image f (E ) is compact in R, whence f (E ) is a bounded set in R. From Theorem 1.1.1, it follows that there exists a sequence fxn g E such
that
n!1

lim f (xn ) = b = sup f (x):

x2E

(1:3:3)

As E is a compact space, by the Bolcano{Weierstrass theorem, from the sequence fxn g one can choose a subsequence fxm g converging to an a 2 E . Since f is a continuous function, by (1.3.3) we have
m!1

lim f (xm ) = f (a) = sup f (x):

x2 E

The proof for the minimum of a function is analogous. Let E and F be metric spaces. A mapping f of E into F is called uniformly continuous if, for an arbitrary " > 0, there exists > 0 such that the inequality d(x0 x00 ) , x0 x00 2 E , yields d(f (x0 ) f (x00 )) ".

Theorem 1.3.10 Every continuous mapping of a compact metric space E into a metric space F is uniformly continuous.
For the proof, see, e.g., Schwartz (1967).

1.4. Maximum function and its properties

15

1.4 Maximum function and its properties

1.4.1 Discrete maximum function
Let U be a topological space and let u ! fi (u) be mappings of U into R, where i 2 I = f1 2 : : : N g. De ne a function ' : U ! R by

u ! '(u) = max fi (u): i2I

(1:4:1)

The function ' is called a maximum function (discrete maximum function).

Theorem 1.4.1 If for every i 2 I , u ! fi(u) is a continuous mapping of U into R, then so is u ! '(u). Proof. Let fung be a sequence converging to u in U . It is clear that max fi (un ) = max fi (u0 ) + (fi (un ) ; fi (u0 )) i2I i2I max fi (u0 ) + max jfi (un ) ; fi (u0 )j: (1.4.2) i2I i2I
By analogy, we get max fi (u0 ) max fi (un ) + max jfi (un ) ; fi (u0 )j: i2I i2I i2I Then, by (1.4.2) and (1.4.3), we have Since un ! u0 in U and fi 's are continuous mappings of U into R, we get
nlim '(un ) = '(u0 ) !1

(1:4:3)

j'(un ) ; '(u0 )j = max fi (un ) ; max fi (u0 ) i2I i2I

max jfi (un ) ; fi (u0 )j: i2I

and so the theorem is proved.

convex on A:

Theorem 1.4.2 Let A be a convex set in a topological space U . If the functions u ! fi (u), i 2 I = f1 2 : : : N g, are convex on A, then the function ' is also

Proof. Let u1 u2 2 A and 2 0 1]. Then fi ( u1 + (1 ; )u2 ) fi (u1 ) + (1 ; )fi (u2 ) '(u1 ) + (1 ; )'(u2 ): Because these inequalities hold for all i 2 I , we have '( u1 + (1 ; )u2 ) '(u1 ) + (1 ; )'(u2 )
which completes the proof.

16

Chapter 1. Basic de nitions and auxiliary statements

Theorem 1.4.3 Let U be a topological space and let u ! fi(u) i 2 I = f1 2 : : : N g, be mappings of U into R that are lower semicontinuous. Let the maximum function be of the form (1.4.1). Then u ! '(u) is a lower semicontinuous mapping of U into R. Proof. Let fung be a sequence converging to u in U . Let us show that, for an
arbitrary " > 0, there exists n" such that

'(un ) '(u) ; " fi (un ) fi (u) ; "

For some k 2 I , we have whence

n n" :

(1:4:4)

Since u ! fi (u) is a lower semicontinuous mapping of U into R, for an arbitrary i 2 I , there exists ni" such that

n ni" :

fk (u) = '(u)

fk (un ) '(u) ; " n nk" : Thus, setting n" = maxi2I ni" , we obtain the inequality (1.4.4), and the theorem
is proved.

1.4.2 General maximum function

Let U and V be metric spaces, and assume that

u v ! F (u v) is a continuous mapping of U V (equipped with the product of the topologies of U and V ) into R.
Further, suppose

(1:4:5)

G is a compact set in V: (1:4:6) By (1.4.5), for an arbitrary xed u 2 U , the partial function v ! F (u v) is a continuous mapping of V into R. As G is compact, by Theorem 1.3.9, for an arbitrary u 2 U there exists a function wu 2 G such that F (u wu ) = max F (u v): v2G Q is a compact set in U: De ne a maximum function ' : Q ! R by u ! '(u) = max F (u v): v2G
Suppose now that (1:4:7) (1:4:8)

1.4. Maximum function and its properties

17

Theorem 1.4.4 Let U , V be metric spaces and let the conditions (1.4.5){(1.4.7) hold. Then, the function u ! '(u) de ned by (1.4.8) is a continuous mapping of Q (equipped with the topology induced by the topology of U ) into R. Proof. Let us show that the following inequality holds: j'(w) ; '(u)j = max F (w v) ; max F (u v) v2G v 2G max jF (w v) ; F (u v)j w u 2 Q: (1.4.9) v 2G
Indeed, we have max F (w v) = max F (u v) + (F (w v) ; F (u v)) v 2G v2G max F (u v) + max jF (w v) ; F (u v)j: v2G v 2G By analogy, we have max F (u v) max F (w v) + max jF (w v) ; F (u v)j: v2G v 2G v 2G Then, (1.4.9) follows from (1.4.10) and (1.4.11). Let fung be a sequence such that (1.4.10) (1:4:11)

un ! u0 in Q: (1:4:12) (1.4.7) and (1.4.12) yield that u0 2 Q. By Theorem 1.3.5, the set Q G equipped with the product of the topologies of Q and G is compact. As Q and G are metric spaces, so is the space Q G. Let d1 and d2 be the metrics of the spaces U and V , respectively, and let = max(d1 d2 ) be the metric on U V . Since Q G is compact, from (1.4.5) and Theorem 1.3.10 it follows that the function u v ! F (u v) is uniformly continuous on Q G (1:4:13) i.e., for an arbitrary " > 0, there exists > 0 such that the inequalities d1 (u0 u00 ) d2 (v0 v00 ) u0 u00 2 Q v0 v00 2 G
yield By (1.4.12), for an arbitrary > 0, there exists N such that, provided n N , the inequality d1 (un u0 ) holds. Then, jF (un v) ; F (u0 v)j " v2G n N : From here, by (1.4.9), we get j'(un ) ; '(u0 )j max jF (un v) ; F (u0 v)j " v2G The theorem is proved.

fung Q

jF (u0 v0 ) ; F (u00 v00 )j ":

n N:

18

Chapter 1. Basic de nitions and auxiliary statements

1.5 Hilbert space

A vector space with a scalar product that is complete with respect to the topology generated by the scalar product is called a Hilbert space. Every Hilbert space is clearly a Banach space. For arbitrary elements x, y of a Hilbert space H , the following inequality| called the Schwartz inequality|holds j(x y)H j kxkH kykH : Theorem 1.5.1 (Riesz) For every Hilbert space H , there exists a unique, oneto-one, linear mapping R of the dual space H onto H possessing the following properties : (Rf x)H = (f x) kRf kH = kf kH f 2 H x 2 H: Notice that here (f x) is the pairing of elements f 2 H and x 2 H , and kf kH = sup j(f x)j:
kxkH

Let H be a vector space over R. A scalar product in H is a bilinear form H H : x y ! (x y)H 2 R that is symmetric, i.e., (x y)H = (y x)H , and positive de nite, i.e., (x x)1 0 and (x x)H = 0 if and only if x = 0. The function H 2 de nes a norm in H . x ! kxkH = (x x)H

1.5.1 Basic de nitions and properties

For the proof, see, e.g., Cea (1971). The operator R 2 L(H H ) is called the Riesz operator for the space H . One can de ne a scalar product in H by setting (f g)H = (Rf Rg)H . Then, the dual of a Hilbert space becomes a Hilbert space, too. The Riesz theorem gives the possibility of identifying the spaces H and H (in that case, a linear functional f 2 H and the element Rf 2 H are identi ed). Theorem 1.5.2 (Lax{Milgram) Let H be a Hilbert space, and let u v ! a(u v) be a bilinear form on H H having the following properties : continuity : there exists a constant c1 > 0 such that ja(u v)j c1 kukH kvkH u v2H coerciveness : there exists a constant c2 > 0 such that a(u u) c2 kuk2 u 2 H: H Then, for every f 2 H , there exists a unique element w 2 H such that a(w v) = (f v) v 2 H: Moreover, the following estimates hold

kwkH c;1 kf kH ? 2

kf kH ? c1 kwkH :

1.5. Hilbert space

19

For the proof, see, e.g., Yosida (1971), Cea (1971). Theorem 1.5.3 Let H be a Hilbert space. If un ! u weakly in H and kunkH ! kukH , then un ! u strongly in H: Proof. From the de nition of a scalar product, it follows that kun ; uk2 = (un ; u un ; u)H = kunk2 ; 2(un u)H + kuk2 : (1:5:1) H H H Since un ! u weakly in H , we have lim (u u) = kuk2 : (1:5:2) H n!1 n H

Taking into account that kunkH ! kukH , by virtue of (1.5.1) and (1.5.2), we conclude that un ! u strongly in H . Two elements u, v of a Hilbert space H are called orthogonal if (u v)H = 0. Two subspaces F1 , F2 of H are called orthogonal if every element of F1 is orthogonal to every element of F2 . Let fui g be a sequence of nonzero orthogonal elements of H , i.e., (ui uj ) = 0 if i 6= j . If the orthogonal system fui g is complete in H (that is, the least closed subspace containing fui g is the whole H ), then this system is called an (orthogonal) basis of H . If, additionally, the norm of each element of the system is equal to 1, then the system fui g is called an orthonormal basis of H . Notice that, in the latter case, (ui uj )H = ij where
ij =
(

1 if i = j 0 if i 6= j

is the Kroneker delta. Let F be a subspace of a Hilbert space H . The set of all the elements of H that are orthogonal to F is called the orthogonal complement of F and is denoted by F ? . Lemma 1.5.1 Let F be a closed subspace of a Hilbert space H . Then, every element u 2 H can be uniquely expanded in the sum u = v + w, where v 2 F and w 2 F ? . The element v is characterized by the following equivalent properties :

ku ; vkH ku ; gkH g 2 F: (u ; v g)H = 0 g 2 F:

For the proof, see, e.g., Cea (1971), Kantorovich and Akilov (1977). Remark 1.5.1 If the hypothesis of Lemma 1.5.1 is satis ed, the element v is called the orthogonal projection of u onto F , and denoted by v = Pu. The operator P is called the orthogonal projection of H onto F Pu = u u2F kP kL(H F ) = 1:

20

Chapter 1. Basic de nitions and auxiliary statements

Remark 1.5.2 Let H be a Hilbert space with a scalar product (f g)H and norm kf kH . Let (f 1g)1 be another symmetric positive de nite form on H H and let kf k1 = (f f )12 . Suppose that there exist positive constants c0 and c1 such that c0 kf kH kf k1 c1 kf kH f2H i.e., the norms k kH and k k1 are equivalent. Denote by H1 the Hilbert space that

coincides as a set with H and the scalar product in which is given by the form ( )1 . Then, evidently, H and H1 coincide as topological spaces, though these are di erent as Hilbert spaces. Despite this, in order to not introduce new notations, we will sometimes say that H is a Hilbert space with the scalar product ( )1 , i.e., we will use the same symbol for di erent spaces in such a case, we will always indicate the form of the scalar product in H .

1.5.2 Compact and selfadjoint operators in a Hilbert space

Let U , V be Hilbert spaces and let J 2 L(U V ), i.e., J is a linear continuous operator acting from U into V . The operator J is called compact if the image J (B ) of every bounded set B U is relatively compact, i.e., the closure J (B ) is a compact set in V . This de nition is equivalent to the following: the image of every weakly convergent sequence from U under the action of the operator J is a sequence that strongly converges in V , i.e., un ! u weakly in U implies Jun ! Ju strongly in V. Let H be a Hilbert space and let A 2 L(H H ). The operator A is called selfadjoint if (Au v)H = (u Av)H

De nitions

u v2H u2H

A is called strictly positive if

(Au u) 0 and the equality to zero takes place only if u = 0. A real number for which the equation

Au = u
has a solution u 6= 0 is called an eigenvalue, and u is called an eigenfunction. The vector subspace generated by all the eigenfunctions belonging to an eigenvalue is a closed subspace of H , which is called the eigenspace of . The dimension of this space is called the multiplicity of .

1.5. Hilbert space

21

Assume H to be an in nite-dimensional Hilbert space. For strictly positive operators, the following theorem holds. Theorem 1.5.4 Let A be a selfadjoint, strictly positive, linear, compact operator acting from H into H . Then, the set of all its eigenvalues i disposed in nonincreasing order, so that every eigenvalue is taken as many times as its multiplicity, forms an in nite sequence of positive numbers convergent to zero : lim = 0: (1:5:3) 1 2 n n!1 n

Spectrum of a selfadjoint operator

There exists a countable orthonormal sequence of eigenfunctions fuig, Aui = i ui (ui uj )H = ij ij 1 (1:5:4) that forms a basis of H . Moreover, (Au u)H (Au1 u1 )H (1:5:5) 1 = sup (u u) = (u u ) : H 1 1H u2H
u6=0

For the proof, see, e.g., Gould (1966). We proceed study the case when the operator A is not positive. Theorem 1.5.5 Let A be a selfadjoint, linear, compact operator acting from H into H . Then, the equation Au = u (1:5:6) has a countable set of eigenvalues f i g in which every eigenvalue is taken as many times as its multiplicity, and every nonzero eigenvalue has a nite multiplicity. Zero is the limit element of the sequence f i g, i.e., limi!1 i = 0. All the eigenvalues of the equation (1.5.6) satisfy the condition m = kvk =1(Av v)H i sup (Av v)H = M: inf (1:5:7)
H

If M 6= 0, then M is an eigenvalue. Moreover, if u is an eigenfunction belonging to M such that kukH = 1, then M = (Au u)H . If m 6= 0, then m is an eigenvalue and if w is an eigenfunction belonging to m such that kwkH = 1, then (Aw w)H = m. For the proof, see, e.g., Riesz and Sz.-Nagy (1972). Let us divide the positive and negative eigenvalues of the operator A into two sequences,
+ + 1 2

kvkH =1

:::

:::

the rst sequence being nonincreasing and the second one nondecreasing, and every eigenvalue appear in these sequences as many times as its multiplicity. Each of these sequences may be nite (and even empty) or in nite.

22 Let

Chapter 1. Basic de nitions and auxiliary statements

u+ u+ : : : 1 2

u ; u; : : : 1 2

be the corresponding sequences of the orthonormal eigenfunctions.

Lemma 1.5.2 Under the assumptions of Theorem 1.5.5, for the n-th positive
eigenvalue, the following representation holds :
+ + + max n = u2Q+ (Au u)H = (Aun un )H n

(1.5.8)

where

Q+ = u j u 2 H kukH = 1 (u u+)H = 0 i = 1 2 : : : n ; 1 n i
and for the n-th negative eigenvalue
; ; n = umin (Au u)H = (Aun un )H 2Q; n ;

where

Q; = u j u 2 H kukH = 1 (u u; )H = 0 i = 1 2 : : : n ; 1 : n i
For the proof, see, e.g., Riesz and Sz.-Nagy (1972). Lemma 1.5.2 allows one to calculate the n-th eigenvalue and eigenfunction, provided known are the eigenfunctions with less numbers. The following lemma is based on the minimax-method, which gives a direct rule of nding the n-th eigenvalue.

Lemma 1.5.3 Let h1 h2 : : : hn;1 be arbitrary (n ; 1) elements of the space H and let = (h1 h2 : : : hn;1 ) be the maximal value of (Au u)H on the set of the elements u 2 H satisfying the condition
kukH = 1
(u hi )H = 0

i = 1 2 : : : n ; 1:

Then, the minimal value of the function on all the possible hi is equal to + and n it is achieved for

hi = u+ i

i = 1 2 : : : n ; 1:

An analogous statement on ; is obtained if we trade the places of \the n minimal value" and \the maximal value." For the proof, see Riesz and Sz.-Nagy (1972). In the sequel, we will need to compare the eigenvalues of di erent operators, so that we will use the following statement.

1.5. Hilbert space

23

Theorem 1.5.6+(Weyl) ; A1 and A be compact selfadjoint operators acting in Let

H . Denote by 1n , + ( 1n , ; ) the n-th positive (negative ) eigenvalues of the n n operators A1 , A, respectively. Then, the following estimates hold :

Proof. Introduce the set Q+n = u j u 2 H kukH = 1 (u u+i )H = 0 i = 1 2 : : : n ; 1 1 1

where u+i is the eigenfunction of the operator A1 belonging to the eigenvalue +i . 1 1 By using Lemmas 1.5.2 and 1.5.3, we get
+

j +; n ;; jn

+j 1n ; 1n j

kA ; A1 k kA ; A1 k

8n 8n:

(1.5.9)

u2Q+n 1 u2Q1n

sup (Au u)H sup (A1 u u)H + sup j(Au u)H ; (A1 u u)H j + +
u2Q1n
+ 1n + + kA ; A + + kA ; A k: 1 1n

Analogously, we get

Thus, the rst estimate in (1.5.9) is proved. Further, introduce the set

1 k:

Q;n = u j u 2 H kukH = 1 (u u;i )H = 0 i = 1 2 : : : n ; 1 1 1 where u;i is the eigenfunction of the operator A1 belonging to the eigenvalue ;i . 1 1
Again, by using Lemmas 1.5.2 and 1.5.3 , we get
;

u2Q;n 1 u2Q;n 1

inf (Au u)H

1n ; kA ; A1 k:
;

inf (A1 u u)H ; sup j(Au u)H ; (A1 u u)H j u2Q;n 1

;

and

So, the second estimate in (1.5.9) is also proved.

1n

; kA ; A1 k:

Equation Au = Bu

Let U , V be two Hilbert spaces such that

U V

the imbedding U ! V is compact:

(1.5.10)

This means that every bounded set in U is relatively compact in V .

24

Chapter 1. Basic de nitions and auxiliary statements

As usual, by ( )U and ( )V we denote the scalar products in U and V if H is a Hilbert space, f 2 H , and u 2 H , then (f u) stands for the dual pairing of f and u. Since, for a xed u 2 U , the form v ! (u v)U is linear and continuous on U , we can write (u v)U = (Au v) Au 2 U (1.5.11) which de nes an operator A 2 L(U U ) called the canonical isometry corresponding to the scalar product. A is evidently the inverse of the Riesz operator de ned in Theorem 1.5.1 . Let us consider the scalar product (u v)V as a bilinear form de ned on U U . Hence, for any u v 2 U , we write (u v)V = (Bu v) Bu 2 U : (1.5.12) Thus, we have de ned an operator B 2 L(U U ): (1.5.13) Let us show that B is a compact mapping of U into U . (1.5.14) Indeed, let un ! u weakly in U , then un ! u strongly in V , and so j(Bun ; Bu v)j = j(un ; u v)V j ckun ; ukV kvkU c = const > 0: Hence, Bun ! Bu strongly in U . Consider the problem on nding (ui i ) 2 U R, ui 6= 0, such that (ui v)U = i (ui v)V v 2 U: (1.5.15) In virtue of (1.5.11) and (1.5.12), this problem is equivalent to the following one: Aui = i Bui : (1.5.16) Denote by A;1 the inverse operator of A, i.e., the Riesz operator. Then, (1.5.16) is equivalent to the relation ui = i A;1 Bui : (1.5.17) Here, stands for the composition of mappings. Since A;1 2 L(U U ), it follows from (1.5.13) and (1.5.14) that A;1 B is a linear compact mapping of U into U . (1.5.18)

1.5. Hilbert space

25

Further, by using (1.5.11) and (1.5.12), we have (A;1 Bu v)U = (Bu v) = (Bv u) = (A;1 Bv u)U = (u A;1 Bv)U u v2U (A;1 Bu u)U = (Bu u) = (u u)V u 2 U: (1.5.19) Thus, A;1 B is a selfadjoint, strictly positive operator. Taking (1.5.18) into account, setting i = 1i , and applying Theorem 1.5.4 to the problem (1.5.17), we obtain the following result. Theorem 1.5.7 Let the condition (1.5.10) be satis ed and let operators A, B be de ned by (1.5.11), (1.5.12). Then, the equation (1.5.16) (equivalently, the equation (1.5.15)) has a countable set of eigenvalues 0< 1 2 lim = 1 i i!1 i
and a corresponding sequence of eigenfunctions fui g that forms an orthogonal basis in U . If U is dense in V , then fui g forms an orthogonal basis in V . Moreover, inf (u u)U (u1 u1 )U i = u2U (u u) = (u u ) : V 1 1V u6=0

Suppose now that U is a Hilbert space, Y is a Banach space, U Y , and the (1.5.20) imbedding U ! Y is compact. Suppose we are given a bilinear, symmetric, continuous form b on Y Y : b(u v) = b(v u) u v2Y (1.5.21) jb(u v)j ckukY kvkY u v 2 Y c = const > 0: (1.5.22) Consider the following eigenvalue problem: (ui i ) 2 U R ui 6= 0 v 2 U: (1.5.23) i (ui v )U = b(ui v ) For a xed element u 2 U , the form v ! b(u v) is linear and continuous on Y , and all the more on U . Therefore, there exists an operator B 2 L(U U ) such that b(u v) = (Bu v) Bu 2 U : (1.5.24) By (1.5.21), the operator B is selfadjoint. By using (1.5.20), analogously to above, we establish that B is a compact mapping of U into U . (1.5.25)

1.5.3 Theorem on continuity of a spectrum

26

Chapter 1. Basic de nitions and auxiliary statements

The problem (1.5.23) is equivalent to the following one: (ui i ) 2 U R ui 6= 0 (1.5.26) i Aui = Bui where A is the canonical isometry corresponding to the scalar product in U (see (1.5.11)). The problem (1.5.26) may be represented in the form (ui i ) 2 U R ui 6= 0 A;1 Bui = i ui : (1.5.27) The operator A;1 B is selfadjoint (see (1.5.19)). Since A;1 2 L(U U ), it follows from (1.5.25) that A;1 B is a compact mapping of U into U . Now, by applying Theorem 1.5.5 to the problem (1.5.27), we get Theorem 1.5.8 Let the conditions (1.5.20) be satis ed and let b be a bilinear form on Y Y for which the formulas (1.5.21), (1.5.22) hold. Then, the problem (1.5.23) has a countable set of eigenvalues f i g, where every eigenvalue is taken as many times as its multiplicity every eigenvalue that is not equal to zero is of nite multiplicity, and the following inequalities hold

m = kuinf=1 b(u u) k
U

kukU =1

sup b(u u) = M:

(1.5.28)

If m 6= 0, then m is an eigenvalue if M 6= 0, then M is an eigenvalue. Moreover, limi!1 i = 0. Note that for proving (1.5.28), one uses the equalities

(A;1 Bu u)U = (Bu u) = b(u u)

u 2 U:
(1.5.29)

Remark 1.5.3 Let X be a Hilbert space and let ' be a bilinear, continuous, symmetric form on X X . Then, the following equality holds
k'kL2 (X X R) = sup j'(u u)j:
kukX =1

For let

Q = sup j'(u u)j:

kukX =1

It follows from the de nition of a norm of a bilinear form (see subsubsec. 1.3.1) that

Q k'kL2(X X R):
Let us show that the inverse inequality holds. Indeed, we have j'(u v)j = 1 ('(u + v u + v) ; '(u ; v u ; v)) 4

1.5. Hilbert space

27

1 Q;ku + vk2 + ku ; vk2 = 1 Q;kuk2 + kvk2 X X X X 4 2 It follows from here that k'kL2 (X X R) = sup j'(u v)j Q
kukX

u v2X

1 kvkX 1

and so (1.5.29) holds. Denote by U1 the set of the bilinear, symmetric, continuous, coercive forms on U U : U1 = a j a 2 L2 (U U R) a(u v) = a(v u) u v 2 U (1.5.30) a(u u) c(a)kuk2 u 2 U c(a) = const > 0 : U Here, c(a) denotes that the positive constant in the inequality of coerciveness depends on an element a. Let further U2 be the set of the bilinear, symmetric, continuous form on Y Y: U2 = b j b 2 L2 (Y Y R) b(u v) = b(v u) u v 2 U : (1.5.31) Suppose now that fang1=1 U1 a0 2 U1 an ! a0 in L2 (U U R) (1.5.32) n 1 fbngn=1 U2 bn ! b0 in L2 (Y Y R) (1.5.33) Consider the following eigenvalue problem: (uin in ) 2 U R uin 6= 0 v 2 U n = 0 1 2 ::: (1.5.34) in an (uin v ) = bn (uin v ) Theorem 1.5.9 Let the conditions (1.5.20) be satis ed and let sets U1, U2 be de ned by (1.5.30), (1.5.31). Let (1.5.32) and (1.5.33) hold. Then, for any n = 0 1 2 : : : , there exists a countable set of eigenvalues of the problem (1.5.34). Let + + ::: ; ; (1.5.35) 1n 2n 1n 2n : : : be the two sequences consisting of the positive and negative eigenvalues ordered so that the rst sequence is nonincreasing and the second one is nondecreasing and every eigenvalue is taken as many times as its multiplicity. Then, lim + = + 8i lim ; = ;0 8j: (1.5.36) i0 j n!1 in n!1 jn
Moreover, the convergence is uniform, i.e., for an arbitrary " > 0, there exists n" such that, for all n n" ,

j j

in ; i0 j ; ; jn ; j 0 j
+ +

" "

8i 8j:

28

Chapter 1. Basic de nitions and auxiliary statements

Remark 1.5.4 Each of the sequences in (1.5.35) may be either empty, or nite, or in nite. To prove Theorem 1.5.9, we need the following lemma. Lemma 1.5.4 Let the conditions (1.5.20) be satis ed, let sets U1, U2 be de ned by the expressions (1.5.30), (1.5.31), and let (1.5.32), (1.5.33) hold true. Let U1 be a subspace of U (in particular, it may be that U1 = U ) and let b b fn = sup an (u u) ; a0 (u u) n = 1 2 ::: 0 (u u) u2U1 n (u u)
u6=0

Then

U U , we get
where

Proof. Since the bilinear form a0(u v) is symmetric, continuous, and coercive on
b b fn = sup an (u u) ; a0 (u u) u2U0 n (u u) 0 (u u) U0 = u j u 2 U1 a0 (u u) = 1 :
n n

lim f = 0: n!1 n

(1.5.37)

(1.5.38)

By (1.5.32) and (1.5.33), jan (u u) ; 1j = jan (u u) ; a0 (u u)j jbn (u u) ; b0 (u u)j

u 2 U0 u 2 U0

Therefore, for an arbitrary " > 0, there exists k" such that bn (u u) b0 (u u) n k" u 2 U0 : (1.5.39) an (u u) ; a0 (u u) " Now, (1.5.37) follows from (1.5.38) and (1.5.39). Proof of Theorem 1.5.9. Since an 2 U1, n = 0 1 2 : : : , for any n one can consider the Hilbert space Vn as the set U in which the scalar product and the norm are generated by the form an . Notice that all the norms generated by an 's are equivalent to each other, and equivalent to the norm of the space U . Hence, (1.5.20) implies that the imbedding of Vn into Y is compact. Now, from Theorem 1.5.8 it follows that there exists a countable set of eigenvalues of the problem (1.5.34). Divide these eigenvalues into the two sequences (1.5.35), and let u+n u+n : : : u;n u;n : : : (1.5.40) 1 2 1 2

lim = 0 n!1 n lim = 0: n!1 n

1.5. Hilbert space

29

be the corresponding sequences of eigenfunctions. The problem (1.5.34) is equivalent to the following one (uin in ) 2 Vn R

uin 6= 0

in uin = A;1 n

Bn uin :

(1.5.41)

Here, A;1 is the inverse of the operator An generated by the scalar product in Vn , n Bn the operator generated by the bilinear form bn ,

an (u v) = (An u v) bn(u v) = (Bn u v)

Analogously to above, we see that

u v 2 Vn u v 2 Vn :

(1.5.42) (1.5.43) (1.5.44) (1.5.45) (1.5.46)

A;1 Bn is a linear, compact, selfadjoint operator in Vn . n

We have

an (A;1 Bn u u) = (Bn u u) = bn (u u) n
+

u 2 Vn n = 0 1 2 : : :

From (1.5.41), (1.5.44), (1.5.45), and Lemma 1.5.2, we have that, for n = 0 1 2 : : : ,

bn (u u) in = sup an (u u) + u2Gin
where

G+ = u j u 2 Vn u 6= 0 an (u u+ ) = 0 k = 1 2 : : : i ; 1 : in kn
Analogously inf bn (u u) jn = u2G; an (u u) jn
;

(1.5.47) (1.5.48) (1.5.49)

G; = u j u 2 Vn u 6= 0 an (u u; ) = 0 k = 1 2 : : : j ; 1 : jn kn

Adding to the set G+ the zero of the space U , we get a subspace G+ . If G+ in in in is considered as a subspace of Vm , m 6= n, then the orthogonal complement of G+ in to the space Vm has dimension i ; 1. Therefore, by using Lemma 1.5.3 and taking (1.5.45) to notice, we get
+

in
+ i0

b sup an (u u) u2G+ n (u u) i0 b sup a0 (u u) u2G+ 0 (u u) in

n = 1 2 ::: n = 1 2 :::

(1.5.50) (1.5.51)

30

Chapter 1. Basic de nitions and auxiliary statements

Now, from (1.5.46) and (1.5.51), we derive

+ i0

b sup a0 (u u) + 0 (u u) u2Gin

b + + sup an (u u) + in = + + in in + n (u u) u2Gin :

(1.5.52)

where

b0 (u u) bn (u u) + in = sup a0 (u u) ; an (u u) + u2Gin
Upon (1.5.46) and (1.5.50),
+

(1.5.53)

in

b sup an (u u) + n (u u) u2Gi0

b + + sup a0(u u) + in = + + in i0 + 0 (u u) u2Gi0 :

(1.5.54)

where

b0(u u) bn(u u) + in = sup a0 (u u) ; an (u u) + u2Gi0

By (1.5.53), (1.5.55), and Lemma 1.5.4 lim =0 n!1 in Hence, (1.5.52) and (1.5.54) give lim = i0 n!1 in
;

(1.5.55)

lim = 0 n!1 in
+ +

8i:

8i:
(1.5.56)

Next, taking into account Lemmas 1.5.2, 1.5.3 and (1.5.45), (1.5.48), we infer
j0

b inf; a0 (u u) u2Gjn 0 (u u)

b ; ; inf; an(u u) ; jn = ; ; jn jn u2Gjn n (u u) :

where

b0 (u u) bn (u u) ; jn = sup a0 (u u) ; an (u u) u2G; jn
;

Analogously, we have
jn

b inf; an (u u) u2Gj0 n (u u)

b ; ; inf; a0 (u u) ; jn = ;0 ; jn j u2Gj0 0 (u u) b0(u u) bn(u u) ; jn = sup a0 (u u) ; an (u u) : ;0 u2Gj

(1.5.57)

1.5. Hilbert space

31

From Lemmma 1.5.4 lim =0 n!1 jn

;

lim =0 n!1 jn
; ; ;

8j:

(1.5.56) and (1.5.57) yield lim = j0 n!1 jn Let

8j:

b b fn = sup an(u u) ; a0(u u) : (u u) 0 (u u) u2U n

u6=0 in ; jn
+

Evidently
+ fn in 8i ; fn jn 8j: Thus, taking to notice that limn!1 fn = 0 (by Lemma 1.5.4), we conclude that, for any " > 0, there exists n" such that, for all n n" , j + ; + j " 8i j ; ; ;0 j " 8j: in i0 jn j

fn fn

The theorem is proved. Remark 1.5.5 Suppose the conditions of Theorem 1.5.9 hold and for n = 0 one of the sets in (1.5.35) is either empty, or nite, and zero is an eigenvalue of the problem (1.5.34) for n = 0. Let, for example, the set of the negative eigenvalues for n = 0 contains k elements, k 0. In accordance with (1.5.35), we set
;

1n

2n

kn

(k+1)n

with ; +1)0 = 0. Then, by using an argument similar to that in the proof of (k Theorem 1.5.9, one can verify that limn!1 ; +1)n = 0. (k From the Riesz theorem it follows that a Hilbert space can be identi ed with its dual. However, such an identi cation is not always suitable. In study of very many problems, one deals with di erent Hilbert spaces that are ordered with respect to imbedding. Then, one of the spaces is identi ed with its dual, and is called the zero space. Usually, as a zero space, one chooses a Hilbert space the scalar product in which is of a rather simple form (for instance, the space L2 ( )). As for the other spaces, one supposes that the pairing of dual elements is determined by the extension of the scalar product of the zero space. The following characterization of dense linear subsets holds.

1.5.4 Imbedding of a Hilbert space in its dual

32

Chapter 1. Basic de nitions and auxiliary statements

Theorem 1.5.10 Let V be a Hilbert space and let D be its linear subset. Then,

D is dense in V if and only if any linear continuous functional on V vanishing on D vanishes on the whole space V:

Proof. The \only if" part is obvious. Let us prove the \if" part. Let D be the closure of D, let P be the orthogonal projection on D, and let P 0 be the transposition of P that is the mapping of V into V de ned by
(P 0 f u) = (f Pu)

f 2 V u 2 V: u 2 D:

Here, as usual, (g v) is the dual pairing of g and v. Evidently, P 0 f ; f 2 V and (P 0 f ; f u) = (f Pu ; u) = 0 Hence, by the hypothesis of the theorem, (P 0 f ; f v) = 0

v 2 V:

Therefore, P 0 is the identical operator on V , and so P is the identity on V and D = V , concluding the proof. Further, let V , H be Hilbert spaces such that

H the imbedding is continuous and dense.

(1.5.58)

Let us choose H as a zero space. Then, the scalar product (u v)H is identi ed with the dual pairing on H H , i.e., for f v 2 H , we have (f v) = (f v)H .

Theorem 1.5.11 Let V , H be Hilbert spaces, let (1.5.58) hold, and let H be a
V H V

zero space. Then, one can identify V and H with dense linear subsets of the dual space V ,

(1.5.59)

where the imbeddings are continuous and the dualization between V and V is identi ed with the unique extension of the scalar product in H:

Proof. To every element u 2 H , one can place in correspondence the linear continuous functional fu in the space V given by
(fu v) = (u v)H

v 2 V:

(1.5.60)

Thus, one de nes the linear continuous mapping u ! fu of the space H into V . This mapping is injective. Indeed, if fu = 0, then by (1.5.60) (u v)H = 0

v2V

1.5. Hilbert space

33

and since V is dense in H , u = 0. So, one can identify the elements u and fu , which gives

H V : Let us show that H is dense in V . To this end, we use Theorem 1.5.10, noticing that the linear form u de ned on V is an element of V , because a Hilbert space is re exive. So, u 2 V = V is a linear form on V vanishing on H . Then (u v) = (u v)H = 0 v 2 H: Hence, u = 0, which yields that H is dense in V . Let now f be an arbitrary element of V . Since H is dense in V , there exists a sequence of elements of H , fung, such that, for the linear forms fn 2 V given
by the relations (fn v) = (un v)H it holds that lim kf ; f kV = 0: n!1 n From this and (1.5.61), we conclude (f v) = nlim (un v)H !1

v2V

(1.5.61)

v2V

i.e., the dualization between V and V is identi ed with the unique extension of the scalar product (u v)H de ned on H V H H , concluding the proof.

1.5.5 Scales of Hilbert spaces and compact imbedding

Compact imbedding of a Hilbert space
Let U , V be Hilbert spaces such that U V , U dense in V , the imbedding of U into V is com(1.5.62) pact. Choose V as a zero space, i.e., identify V with its dual V . Consider the following eigenvalue problem: (uj v)U = j (uj v)V v 2 U: (1.5.63) From (1.5.62) and Theorem 1.5.7 we deduce the existence of a countable set of eigenvalues f j g such that 0<
1 2

lim = 1 j !1 j

(1.5.64)

34

Chapter 1. Basic de nitions and auxiliary statements

and of a corresponding sequence of eigenfunctions fuj g which solve the problem (1.5.63) and form an orthonormal basis in V and orthogonal in U . Thus, (ui uj )V = ij (ui uj )U = i ij : (1.5.65) Every element u 2 V has a representation

u=
By (1.5.65)

1 X

i=1

(u ui )V ui :

(1.5.66)

kuk2 = V
and if u 2 U , then

1 X

i=1

(u ui)2 < 1 V
i (u ui )2 V

(1.5.67)

kuk2 = U

1 X

fui g is a basis in V and U , every element g 2 U has a representation

g=
Let us show that
1 X

From (1.5.62) and Theorem 1.5.11, it follows that V is dense in U , and since (g ui )ui :
;1

i=1

< 1:

(1.5.68)

i=1

(1.5.69)

kgk2 = U

1 X

Indeed, in virtue of (1.5.66), (1.5.68), and (1.5.69), we get, for g 2 U and any u 2 U,

i=1

2 i (g ui ) :

(1.5.70)

j(g u)j =
=

1 X

i=1 1 X

(g ui )(u ui )V
;1 2 ;1

i=1 1 X i=1

i (g ui )

2 i (u ui )V 1 2

i (g i

u )2

kukU :
(1.5.71)

Hence,

kgkU

1 X

;1

i=1

i (g i

1 2 2: u)

1.5. Hilbert space

35

Let v be of the form

v=

n X i=1

;1

i (g ui )ui :

(1.5.72)

Evidently, v 2 U and by (1.5.66), (1.5.68) we have

kvk2 = U j(g v)j =

From here
n X i=1
;1

n X i=1

;1

2 i (g ui ) :

(1.5.73)
1 2

Upon (1.5.69), (1.5.72), and (1.5.73), we infer

i (g i = n X i=1

u )2

n X i=1
1 2

;1

i (g i

u )2

kvkU :

;1

i (g i

u )2

kgkU : kgkU :

The latter inequality holds for an arbitrary n, and therefore

1 X ;1

i=1

i (g i

u )2

1 2

This together with (1.5.71) yields (1.5.70). Let the conditions (1.5.62) be satis ed and let H , ;1 spaces distinguished by

Scales of Hilbert spaces

1, be the Hilbert (1.5.74)

H = uju2U
and endowed with the norm

1 X

i=1
1 X

2 i (u ui ) < 1
1 2

kukH =

i=1

2 i (u u i ) :

(1.5.75)

From (1.5.67), (1.5.68), and (1.5.70), it follows that H0 = V = V H1 = U H;1 = U and

H = H;

> 0:

36

Chapter 1. Basic de nitions and auxiliary statements

Theorem 1.5.12 Let U , V be two Hilbert spaces for which (1.5.62) holds and let spaces H , ;1 1, be de ned by (1.5.74), (1.5.75). Then, the imbedding of H 1 into H 2 , where ;1 2 < 1 1, is compact. Proof. Let fvng H 1 and vn ! 0 weakly in H 1 : (1.5.76) Let us show that then vn ! 0 strongly in H 2 : (1.5.77) From (1.5.75) and (1.5.76) we get
kvn k2 1 = H
By (1.5.75),
1 X

i=1

2 1 i (vn ui )

const

8n:

(1.5.78)

kvn k2 2 = H

1 X

i=1 M X i=1

2 2 i (vn ui )

2 (v

i ! 1 and

Let now " be an arbitrary positive number. Taking to notice that i ! 1 as 1 > 2 , we infer from (1.5.78) that there exists M such that
1 X

1 X 2 2; 1 1 n ui )2 + M +1 i (vn ui ) i=M +1

(1.5.79)

By (1.5.76), there is N such that

M X i=1
2 2 i (vn ui )

i=M +1

M +1

2; 1

2 1 i (vn ui )

"

8n:

(1.5.80)

"

n N:

(1.5.81)

Finally, (1.5.77) follows from (1.5.79){(1.5.81), concluding the proof.

1.6 Functional spaces that are used in the investigation of boundary value and optimal control problems
1.6.1 Spaces of continuously di erentiable functions
Let be a domain in Rn , i.e., an open connected set in Rn , x = (x1 x2 : : : xn ) 2 . Further, let k = (k1 k2 : : : kn ) be an ordered row of nonnegative integers ki ,

1.6. Functional spaces

:::

37

which will be referred to as a multi-index. To every multi-index k we place in correspondence the di erential operator The nonnegative integer
k1 + +kn : Dk = @k1 @x1 @xkn n

A real-valued function u determined in is said to belong to the space C l ( ) if it has continuous derivatives in up to and including order l. This means that every derivative of order jkj l exists at all points of and coincides in these points with a continuous function on . The norm in C l ( ) is de ned by the formula

jkj = k1 + + kn is called the order of the operator Dk . If jkj = 0, then Dk f = f .

kukC l( ) =

jkj l x2

sup Dk u(x) :

For l = 0, we get the space of continuous functions on , which is denoted by C ( ). If is a bounded open set in Rn , then C l ( ) is a Banach space, and, in the de nition of the norm, sup may be changed by max. By C 1 ( ) we denote the space of in nitely di erentiable functions in .

1.6.2 Spaces of integrable functions

The space of (classes of) real-valued functions f which are measurable on such that 1 Z p dx p < 1 1 p<1 kf kLp( ) = jf (x)j

and

will be referred to as Lp ( ). The space Lp ( ) is a separable Banach space. As long as 1 < p < 1, the spaces Lp ( ) are re exive. The space L2( ) is a Hilbert space equipped with the scalar product Z (u v) = u(x) v(x) dx: A measurable function u : ! R is called essentially bounded if it is equivalent to some bounded function, i.e., if there exists a number M such that ju(x)j M for almost all x 2 . The precise lower bound of such constants is denoted by vrai2max ju(x)j. x The space of (classes of) measurable, essentially bounded functions will be referred to as L1 ( ), which is a Banach space with respect to the norm kukL1( ) = vrai2max ju(x)j: x

38

Chapter 1. Basic de nitions and auxiliary statements

L1 ( ) is called the space of integrable functions. If is a bounded domain in Rn , then Lp ( ) L1 ( ) for all p 2 (1 1].

Theorem 1.6.1 (Lebesgue) Let fuk g be a sequence of integrable functions on which converges almost everywhere (a.e.) on to a function u, i.e., for almost all x 2 the numerical sequence fuk (x)g tends to u(x) as k ! 1. Assume that
there exists an integrable function v on such that

juk (x)j v(x)

Then, the function u is integrable on
k!1

8k a.e. on :
and
Z

lim

uk dx =

u dx:

Let x ! f (x) be an arbitrary continuous function on . The support of f (denote by supp f ) is the least closed set in outside of which f (x) vanishes, i.e., supp f = x j f (x) 6= 0 \ : By D( ) we denote the space of real-valued functions u which are in nitely di erentiable in the domain and have compact supports in The space D( ) is endowed with the inductive limit topology (see, e.g., Schwartz (1966), Rudin (1973)). A sequence fuig D( ) converges to a function u 2 D( ) as i ! 1 if the following conditions hold: 1. there exists a compact set B ui are contained in B such that the supports of all the functions

1.6.3 Test and generalized functions

2. the derivatives of an arbitrary order m 1 of the functions ui uniformly converge to the corresponding derivatives of the function u as i ! 1. By D ( ) we denote the dual of D( ) (the space of distributions or generalized functions), i.e., the space of linear functionals on D( ) that are continuous with respect to the topology of D( ). If T 2 D ( ) and u 2 D( ), then the value of T at u is denoted by (T u) and called the dual pairing between T and u. The space D ( ) is equipped with the strong dual topology, that is, with the topology of uniform convergence on every bounded set in D( ). Notice that the convergence in the weak topology of the space D ( ) is equivalent to the convergence in the strong topology. Thus, a sequence of distributions fTig is said to converge to a distribution T if, for every function u 2 D( ), (Ti u) ! (T u):

1.6. Functional spaces

:::

39

@T For T 2 D ( ), the derivative @xi is de ned by the equality

@T u = ; T @u @xi @xi
which determines the linear continuous mapping

u 2 D( )

@T T ! @x

of D ( ) into D ( ). Placing in correspondence to every function f 2 Lp ( ) the distribution (generalized function) f~ given by the formula

u ! (f~ u) =

f (x) u(x) dx

u 2 D( )

~ we obtain the linear, continuous, injective mapping f ! f of the space Lp ( ) into ~ can be identi ed. Thus, we have D ( ). Then, f and f

D( ) Lp ( ) D ( ):
; k Df ; u = (;1)jkj f Dk u

(1:6:1)

@f Taking (1.6.1) into account, one derives the distribution derivatives @xi of any function f 2 Lp ( ). In general, for f 2 D ( ), we set

u 2 D ( ):

(1:6:2)

According to (1.6.2), a function f 2 Lp( ) is said to has the distribution derivative Dk f which belongs to Lp ( ) if there exists a function wk 2 Lp ( ) such that ; (wk u) = (;1)jkj f Dk u u 2 D( ): The function wk is called the distribution derivative of f 2 Lp ( ), and the notation wk = Dk f is used just as for the ordinary derivative.

1.6.4 Sobolev spaces

De nitions. The imbedding theorem For p 1 and m 2 N (N denoting the set of natural numbers), the Sobolev space Wpm ( ) is de ned as the set all v 2 Lp ( ) whose all distribution derivatives Dk v, jkj m, belong to Lp ( ), and this set is endowed with the norm
kvkWpm ( ) =
X
jkj

Dk v p p ( L

1 p

(1:6:3)

40

Chapter 1. Basic de nitions and auxiliary statements

As the di erential operator Dk maps continuously D ( ) into itself, Wpm ( ) is a Banach space, and it is re exive for 1 < p < 1. When m = 0, the space Wpm ( ) is identi ed with the space Lp ( ), and, for p = 2, W2m ( ) is a Hilbert space. For domains with Lipschitz continuous boundaries, the set of functions C 1 ( ) is dense in Wpm ( ). By W m ( ) we denote the closure of D( ) in the norm of the space Wpm ( ). p The imbedding of a space Wpm ( ) into a space Wpm1 ( 1 ), 1 , is the 1 operator which maps every function from Wpm ( ) into itself, being considered as an element of Wpm1 ( 1 ). There are a number of results known as imbedding theorems. 1 Let us present one of them, see, Sobolev (1963), Adams (1975), Brezis (1983).
a Lipschitz continuous boundary. Then, the imbedding of Wpm ( ) into Lq (Sr ), where Sr is the intersection of with an r-dimensional plane, in particular, Sn = pr , is a bounded operator if n > mp, r > n ; mp, q n;mp , and is a compact pr . For n = mp, it is a compact operator for every nite q . operator if q < n;mp For n < mp, any function u 2 Wpm ( ) is continuous in , and the operator of imbedding of Wpm ( ) into C ( ) is bounded and compact. If n < mp and 0 < l < m ; n , the operator of imbedding of Wpm ( ) into C l ( ) is bounded and compact. p continuous boundary. Then, there exists a continuation operator P such that

Theorem 1.6.2 (imbedding theorem) Let be a bounded domain in Rn with

Theorem 1.6.3 (Calderon) Let be a bounded domain in Rn with a Lipschitz

P 2 L Wpm ( ) Wpm (Rn ) Pu = u a.e. in :
;

For the proof see, Calderon (1961) and Fikhtengolts (1966) for the case n = 2.

Averaging (regularization) of functions

Let !( ) be a nonnegative, in nitely di erentiable function on

R+ = t j t 2 R t 0
which vanishes for 1. For example,

!( ) = exp
0

2 ;1

if 0 if 1

<1 < 1:
,

; If x 2 Rn , then the function x ! ! jxj vanishes for jxj

jxj =

n 2 X 2 1 xi i=1

1.6. Functional spaces

:::

41

and

Z
Rn

! jxj dx = n
(1:6:4)

where is a constant. As an averaging kernel we take

! (x) = 1 n ! jxj :

If f is a locally integrable function in Rn , then the averaging of this function has the form Z f (x) = ! (x ; y) f (y) dy (1:6:5) (actually, the integral is taken over the ball jx ; yj , not over the whole Rn ). If f is de ned only in the domain , then f is de ned in the domain the boundary of which is remote from the boundary of on distance , and ; f 2 C1 . The averaging has the following properties (see Sobolev (1963)). 1. Let f 2 Lp ( ), p 1. Set f (x) = 0 for x 62 . Then, the function f is well-de ned in Rn , in nitely di erentiable, and f ! f in Lp ( ) as ! 0. 2. Let f 2 Lp ( ), p 1. Set f (x) = 0 for x 62 . Then, for all > 0, the function f 7! f is a linear continuous mapping of Lp ( ) into C m ( ) for an arbitrary integer m > 0. 3. Let K be an arbitrary compact set in and let f 2 Wpm ( ). Then, Dk f = (Dk f ) in K as long as jkj m and is su ciently small. Moreover, Dk f ! Dk f in Lp(K ) as ! 0.
Rn

Remark 1.6.1 Property 3 and Theorem 1.6.3 imply that, if is a bounded domain in Rn with a Lipschitz continuous boundary, then every function u 2 Wpm ( ) can be extended to a function u = Pu 2 Wpm (Rn ), and moreover u ! u in Wpm ( ) as ! 0. Theorem on equivalent norms
We will need the following statement.

Wpm ( ), where g(x) =

jk j

Theorem 1.6.4 Let lj , j = 1 2 : : : N , be linear continuous functionals in is a bounded domain in Rn and 1 < p < 1. Assume that,
for each polynomial
X

m;1

ck xk1 1

xkn n

jkj = k1 +

+ kn ck = const

(1:6:6)

42

Chapter 1. Basic de nitions and auxiliary statements

of degree m ; 1, the condition lj (g) = 0, j = 1 2 : : : N , yields that g = 0. Then, the norm in Wpm ( ) de ned by the formula
1 N X p p kv p kv k = D Lp( ) + jlj (v)j j =1 jk j=m

(1:6:7)

is equivalent to the norm (1.6.3).

Proof. Let us establish the inequality kvkWpm( ) ckvk

v 2 Wpm ( )

(1:6:8)

where c is a constant and the norms on the left and right hand sides are de ned by the relations (1.6.3) and (1.6.7). Indeed, if (1.6.8) is not valid, then there exists a sequence fvi g Wpm ( ) such that
X
jkj=m

kvi kWpm ( ) = 1 Dk vi p p ( ) ! 0 L
as i ! 1

8i

as i ! 1

(1.6.9) (1.6.10) (1.6.11)

lj (vi ) ! 0 v ! v0 Dk v ! Dk v0 Dk v ! Dk v0 Dk v0 = 0

j = 1 2 : : : N:

By virtue of (1.6.9) and Theorem 1.6.2, from the sequence fvi g one can choose a subsequence fv g such that weakly in Wpm ( ) weakly in Lp ( ) jkj = m strongly in Lp ( ) jkj m ; 1: a.e. in (1.6.12) (1.6.13) (1.6.14) (1:6:15)

(1.6.10) and (1.6.13) imply that

jkj = m:

Let K be an arbitrary compact set in . For su ciently small > 0, in K we can de ne the function v0 which is the averaging of the function v0 . (1.6.15) and Property 3 of the averaging imply that

Dk v0 = Dk v0 = 0

in K jkj = m:

From here, taking into account the in nite di erentiability of the function v0 in K , we get

v0 =

jkj

m;1

ak xk1 1

xkn n

in K jkj = k1 +

+ kn

(1:6:16)

1.6. Functional spaces

:::

43

ak being constants. By Property 3 of the averaging, we have v0 ! v0 in Lp (K ) as ! 0. Thus, (1.6.16) and the closedness of a nite-dimensional subspace of Lp (K ) yield v0 =
X
jkj

m;1

ak xk1 1

xkn n

in K jkj = k1 +

+ kn :

(1:6:17)

Let K1 be a compact neighborhood of K , and let K1 . Similarly to the above argument, we conclude that the function v0 is de ned by the relation (1.6.17) in K 1 and, hence, in K . There exists a sequence of compact sets fK g such that S K , K K +1 for all , and = 1 K . Since in every K the function =1 v0 has the form (1.6.17), it is determined by this formula in . (1.6.11) and (1.6.12) imply lj (v0 ) = 0 j = 1 2 : : : N: Since v0 is a polynomial in of degree m ; 1, the hypothesis of the theorem yields v0 = 0: (1:6:18) By (1.6.10), (1.6.14), and (1.6.15), we have v ! v0 strongly in Wpm ( ): This makes a contradiction with (1.6.9) and (1.6.18), so that (1.6.8) holds. The inequality inverse of (1.6.8) is implied by the boundedness of the functionals lj . The theorem is proved. The following statement is a consequence of the proof of Theorem 1.6.4. Corollary 1.6.1 Let be a bounded domain in Rn and let 1 < p < 1. Assume that E is a subspace of Wpm ( ) such that the condition that g 2 E is a polynomial of degree m ; 1 yields g = 0. Then, in the space E , a norm de ned by the relation

kv k =
is equivalent to the norm (1.6.3).

X
jkj=m

1 p kv p D Lp( )

In the case p = 2, one uses the notation H m ( ) = W2m ( ): If the boundary S of the domain is regular, more precisely, if S is a manifold of dimension n ; 1 and of the C 1 class, and is placed at one side of S ,

Spaces H s( )

44

Chapter 1. Basic de nitions and auxiliary statements

then one can de ne the spaces H s ( ) for any real s 0 as intermediate spaces between H m ( ), m 2 N , and H 0 ( ) = L2 ( ) (see Lions and Magenes (1972) and subsec. 1.5.5). In case of a smooth boundary, this de nition of the space H s ( ) is independent of a choice of m H s ( ) is a Hilbert space. In particular, if is a bounded regular domain,

H s( ) = u j u 2 H 0( )

1 X

where 2 0 1], i and ui are the eigenvalues and eigenfunctions of the problem (ui v)H m ( ) = i (ui v)H 0 ( ) v 2 H m( ) and (ui uj )H 0 ( ) = ij . The norm in H s ( ) can be de ned by the formula

i=1

2 i (u ui )H 0 ( ) < 1

m=s

kukH s ( ) =

1 X

s By H0 ( ) we denote the closure of D( ) in H s ( ), and by H ;s ( ) the dual space s ( ), s > 0, i.e., (H s ( )) = H ;s ( ). of H0 0 By using a system of local maps, one introduces the notion of a trace of a function as an element of the space H t (S ), t > 0, (H t (S )) = H ;t(S ) for t 0. The following theorem on the trace space holds (see Lions and Magenes (1972)). Theorem 1.6.5 The mapping i ; C 1 ( ) 3 u ! @ u S i = 0 1 2 : : : m ; 1 2 C 1 (S ) m i @
i

i=1

i (u i H ( )

u )2 0

1 2

m = s:

(1:6:19)

where @ u S is the normal derivative of order i on S , can be extended by continuity @ i Q 2 to a linear continuous mapping of H m ( ) into m;1 H m;i; 1 (S ): i=0 m The mapping (1.6.19) is surjective and its kernel coincides with H0 ( ) = W m ( ): 2

Remark 1.6.2 If is a bounded domain with a Lipschitz continuous boundary, then the trace ujS of a function f 2 H m ( ), m 1, on S belongs to the space L2 (S ) and the corresponding imbedding operator u ! ujS is compact, see, e.g.,
Ladyzhenskaya and Uraltseva (1973). Remark 1.6.3 If the boundary S of a domain is not smooth, but consists of a nite number of smooth, open (in S ), and disjoint subsets Si such that

S=

i=1

Si

1.7. Inequalities of coerciveness

45

then to any set Si one can apply Theorem 1.6.5. More precisely, if Si is an (n ; 1)dimensional manifold of the C 1 class, is placed at one side of S , and Si0 is an open set in Si such that Si0 Si , then
Q 1 is a continuous, surjective mapping of H m ( ) onto m;1 H m;i; 2 (Si0 ). i=0

i u ! @ u S0 i = 0 1 2 : : : m ; 1 @ i i

1.7 Inequalities of coerciveness

Let W = m W2lr ( ) be a topological product of spaces W2lr ( ), lr 1, r = r=1 1 2 : : : m, i.e., W is the space of vector functions v = (v1 v2 : : : vm ) de ned on Rn and taking values in Rm , where vr 2 W2lr ( ). The norm in W is de ned through the formula m X kvk2 = kvr k2 2lr ( ) : (1:7:1) W W Further, let Ni , i = 1 2 : : : , be the linear continuous mappings of W into L2 ( ) de ned by the formula
r=1 m X X

1.7.1 Coercive systems of operators

Q

v ! Ni v =
where

r=1 jkj lr

girk Dk vr

(1:7:2)

girk 2 L1 ( ): (1:7:3) The system of operators fNi gi=1 is called W -coercive with respect to (L2 ( ))m if there exists a constant c > 0 such that
X

i=1

kNi vk2 2 ( ) + L

m X r=1

kvr k2 2 ( L

ckvk2 W

v 2 W:

(1:7:4)

Let V be a closed subspace of W (in particular, we can take V = W ). The system of operators fNi gi=1 is called coercive in V if there exists a constant c > 0 such that X kNi vk2 2 ( ) ckvk2 v 2 V: (1:7:5) W L For an arbitrary = ( 1 : : : n ) 2 C n (C being the eld of complex numbers), we set X k k Nir (x ) = girk (x) 1 1 nn : (1:7:6)
jkj=lr

i=1

46

Chapter 1. Basic de nitions and auxiliary statements

An open set

Rn is said to satisfy the cone condition if x + U (e(x) H ) x2

where U (e(x) H ) is a straight circular cone with vertex at the origin, of a xed angle, and of a xed height H , 0 < H < 1, whose axis has direction e(x) depending on x. In particular, a Lipschitz domain satis es the cone condition. The following statement is valid (Besov et al. (1975), Hlavacek and Necas (1970)). Theorem 1.7.1 Let be a bounded domain in Rn satisfying the cone condition. The system of di erential operators

v ! Ni v =

m X X

r=1 jkj lr

girk Dk vr

i = 1 2 :::

where girk 2 L1( ) if jkj < lr and girk 2 C ( ) if jkj = lr , is W -coercive with respect to (L2 ( ))m if (and, in the case when girk are constants for jkj = lr , only if ) the rank of the matrix fNir (x )g is equal to m for an arbitrary complex 2 C n , 6= 0, and any x 2 . In the sequel, we will need also the following theorem. Theorem 1.7.2 Let fNigi=1 be a system of operators that is W -coercive with respect to (L2 ( ))m and let u v ! a(u v) be a bilinear, symmetric, continuous form on W W such that a(v v) 0 v2W (1:7:7) and the conditions

w2W
X

imply w = 0. Then, there exists a positive constant c0 such that

i=1

i=1

kNi wk2 2 ( ) + a(w w) = 0 L

v 2 W: v 2 W:

(1:7:8)

kNivk2 2 ( ) + a(v v) c0 kvk2 W L

(1:7:9)

Proof. Let us show rst that there exists a positive constant c1 such that
X

Here

i=1

kNivk2 2 ( ) + a(v v) c1 kvk2L2 ( L ( kvk2L2 ( (

))m

))m

(1:7:10)

m X r=1

kvr k2 2 ( ) : L

1.7. Inequalities of coerciveness

47

fv(n) g of elements of W such that kv(n) k(L2 (

X

Assume that the inequality (1.7.10) is not valid. Then, there exists a sequence
))m

=1

(1.7.11) (1.7.12)

i=1

kNi v(n) k2 2( ) ! 0 L

a v(n) v(n) ! 0:

Upon (1.7.4), (1.7.11), and (1.7.12), we conclude that the sequence fv(n) g is bounded in W . Let us choose from it a subsequence fv(k) g such that

v(k) ! w v(k) ! w

weakly in W strongly in (L2( ))m :

(1.7.13) (1.7.14) (1:7:15)

The formulas (1.7.11) and (1.7.14) yield

kwk(L2 (
(1.7.13) implies

))m

= 1:

Ni v(k) ! Ni w
X

weakly in L2( ) as k ! 1
X

i = 1 2 ::: :
X

Combining this with (1.7.12) and taking into consideration Theorem 1.2.4, we obtain lim k!1
i=1

kNi v(k) k2 2( ) = lim inf L k!1

i=1

kNi v(k) k2 2( L

i=1

kNi wk2 2 ( ) = 0: L
(1:7:16)

One can easily derive the equality

a(v(k) v(k) ) = a(w w) + 2a(w v(k) ; w) + a(v(k) ; w v(k) ; w):

Then, (1.7.7), (1.7.12), and (1.7.13) yield
k!1

lim a(v(k) v(k) ) = lim inf a(v(k) v(k) ) a(w w) = 0: k!1

(1:7:17)

By virtue of (1.7.16) and (1.7.17), the element w satis es the conditions (1.7.8), so that w = 0. But this makes a contradiction to the equality (1.7.15), hence (1.7.10) is valid. The inequality (1.7.10) implies that
X

i=1

kNi vk2 2 ( ) + a(v v) c2 L

i=1

kNi vk2 2 ( ) + kvk2L2 ( L (

))m

c2 being a positive constant.

v2W (1:7:18)

Now, from (1.7.4) and (1.7.18) we deduce the inequality (1.7.9) with the positive constant c0 = cc2 .

48

Chapter 1. Basic de nitions and auxiliary statements

Remark 1.7.1 Assume that the conditions of Theorem 1.7.2 are satis ed and that V is a closed subspace of W de ned through the relation
V = u j u 2 W a(u u) = 0 :
Then, the relation (1.7.9) implies the inequality
X

i.e., the system of operators fNi g is coercive in V .

i=1

kNivk2 2 ( L

c0 kvk2 W

v2V

Theorem 1.7.3 Let fNigi=1 be a system of operators that is W -coercive with respect to (L2 ( ))m and let V be a closed subspace of W such that the condition
w2V
X X

yields w = 0. Then, there exists a positive number c such that

i=1

i=1

kNi wk2 2 ( ) = 0 L
ckvk2 W v2V

kNi vk2 2 ( L

i.e., the system of operators fNi gi=1 is coercive in V:

Proof. We will use Theorem 1.7.2. Since V m

u v ! a(u v) = 0

W , fNi gi=1 is V -coercive with respect to (L2 ( )) . Choose the zero form for the bilinear, symmetric, continuous, nonnegative form a on W W , i.e., u v 2 W:
Now, setting W = V in Theorem 1.7.2, we get Theorem 1.7.3 as a consequence of Theorem 1.7.2.

We will now obtain Korn's inequality for a two-dimensional space as an example of application of Theorem 1.7.1 we will need this inequality below. Let be 2 a ; bounded domain in R with a Lipschitz continuous boundary, and let W = 1( ) 2. W2 Introduce a system of linear continuous operators Ni mapping W into L2 ( ) through the following formulas:

1.7.2 Korn's inequality

@v N1 v = 2 @x1 1

v = (v1 v2 ) 2 W p @v @v N2 v = 2 @x1 + @x2 2 1

@v N3 v = 2 @x2 : (1.7.19)
2

1.7. Inequalities of coerciveness

49

Comparing (1.7.2) and (1.7.19) we see that girk are constants in the present setting, so that the matrix (Nir (x )) does not depend on x, and due to (1.7.6) it has the form 2 3 2 p 1 p0 5 (Nir (x )) = 4 2 2 2 1 : (1:7:20) 0 22 It is easy to see that, for any 2 C 2 6= 0, the columns of the matrix (1.7.20) are linearly independent, hence the rank of this matrix is equal to 2. Applying Theorem 1.7.1 and taking into consideration (1.7.19), we get
Z X 2

i j =1

@vi + @vj @xj @xi

dx+

2 X

i=1

kvi k2 2 ( L

2 X

i=1

kvi k2 21 ( W

v 2 W (1:7:21)

(1.7.21) is called Korn's inequality. From the physical point of view, the expressions Ni v de ned by the formulas (1.7.19) determine up to constant multipliers the components of the strain tensor of a continuum,

c being a positive constant.

@v "ij (v) = 1 @xi + @vj 2 j @xi caused by a vector function of displacements v = (v1 v2 ).

(1:7:22)

To apply Theorem 1.7.2, we must nd the intersection of the kernel spaces of the operators Ni de ned through the formulas (1.7.19), i.e., we must nd the subspace of functions v 2 W such that

@v @v "ij (v) = 1 @xi + @xj = 0 2 j i

i j = 1 2:

(1:7:23)

From the physical point of view, these functions de ne \in nitesimal rigid displacements of continuum." We denote the subspace of these functions by Q,

@v @v Q = v j v 2 W @xi + @xj = 0 i j = 1 2 : j i Q = v j v = (v1 v2 ) v1 = a1 + a3 x2 v2 = a2 ; a3 x1 ai 2 R :

Z X 2

(1:7:24) (1:7:25)

The space Q is known to be of the form (see, e.g., Hlavacek and Necas (1970)) Let S be the boundary of and let S1 be an open set in S . De ne a bilinear symmetric form on W W by

a(u v) =

S1 i=1

ui vi ds:

50

Chapter 1. Basic de nitions and auxiliary statements

2 X

Taking to notice Remark 1.6.2, we obtain

ja(u v)j

Hence, the form u v ! a(u v) is continuous on W W . Further, let us show that, if v 2 Q and then v = 0. Indeed, suppose that v 2 Q. In view of (1.7.25), the function x ! v(x) has the following representation v(x) = a + Ax (1:7:27) where v = v1 a = a1 A = ;0 a03 : (1:7:28) v a a
2 2 3

i=1

kui kL2(S1 ) kvi kL2(S1 ) ckukW kvkW :

Z ; 2 2 v1 + v2 S1

ds = 0

(1:7:26)

Since S1 is an open set in S , there exist two di erent points x(1) , x(2) belonging to S1 . By virtue of (1.7.26){(1.7.28), we get A(x(1) ; x(2) ) = 0 i.e., the determinant of the matrix A should be equal to zero. So, (1.7.28) implies a3 = 0, and from (1.7.26), (1.7.27) we deduce that a1 = a2 = 0. Now, from (1.7.21) and Theorem 1.7.2 we derive the inequality
Z X 2

i j =1

@vi + @vj 2 dx + Z ;v2 + v2 ds ckvk2 1 2 W @xj @xi S1 v 2 W c = const > 0: V = uju 2 W

Z ; u2 + u2 1 2 S1

(1.7.29)

Introduce the notation

ds = 0 :

(1:7:30)

V is obviously a subspace of W . The set V is closed in W since it is the prototype of the closed set f0g under the continuous mapping u!
Z X 2

of W into R. (1.7.29) and (1.7.30) yield

i j =1

Z ; u2 + u2 1 2 S1
2

ds v 2 V:
(1:7:31)

@vi @vj @xj + @xi

dx ckvk2 W

1.8. Theorem on the continuity of solutions of functional equations

51

1.8 Theorem on the continuity of solutions of functional equations

Let F , G be normed vector spaces and let A be a linear continuous operator acting from F into G, i.e., A 2 L(F G). Remind that a mapping is called invertible if it is a bijection and the inverse bijection A;1 is also continuous, i.e., A;1 2 L(G F ).

Theorem 1.8.1 Let Y be a topological space, let F , G be normed vector spaces, and let U be the set of the invertible elements of the space L(F G) equipped with the topology induced by the topology of L(F G). Assume that B : h ! B (h) is a continuous mapping of Y into U . De ne a function Q : (h f ) ! uh f , where (h f ) 2 Y G and uh f is the solution to the problem uh f 2 F B (h)uh f = f: (1:8:1)
Then, Q is a continuous mapping of Y

G into F . Proof. Let fhn fng be a sequence of elements of Y G such that hn ! h0 in Y fn ! f0 in G:

The rst relation in (1.8.2) implies that B (hn ) ! B (h0 ) in U :

(1:8:2) (1:8:3)

Denote the set of the invertible elements of the space L(G F ) by U ;1 . Then the bijection A ! A;1 of U onto U ;1 is a homeomorphism (cf. Schwartz (1967)). Hence, (1.8.3) yields (B (hn ));1 ! (B (h0 ));1 in U ;1 : (1:8:4) Here, (B (hn ));1 (B (h0 ));1 are the bijections inverse of B (hn ) and B (h0 ), and the set U ;1 is equipped with the topology induced by the topology of the space L(G F ). Let ; un = uhn fn Bn = B (hn ) Bn 1 = (B (hn ));1 n = 0 1 2 : : : (1:8:5) Then Therefore
; un = Bn 1 fn ; u0 = B0 1 f0 :

pleting the proof.

; ; ; kun ; u0 kF = kBn 1 (fn ; f0 ) + (Bn 1 ; B0 1 )f10 kF ;1 ;1 k ; kBn L(G F )kfn ; f0 kG + kBn ; B0 kL(G F )kf0kG : This inequality together with (1.8.2), (1.8.4), (1.8.5) implies un ! u0 in F , com-

52

Chapter 1. Basic de nitions and auxiliary statements

1.9 Di erentiation in Banach spaces and the implicit function theorem

1.9.1 Frechet derivative and its properties
Let f be a mapping of an open set U in a Banach space E into a Banach space F . The mapping f is said to be Frechet di erentiable at a point a 2 U if there exists a linear continuous mapping L 2 L(E F ) such that, for any a + h 2 U , the equality f (a + h) = f (a) + Lh + '(h)khkE holds, where '(h) 2 F and k'(h)kF ! 0 as h ! 0 in E and h 6= 0. The mapping L is called the Frechet derivative of the mapping f at the point 0 a, and it is denoted by L = f 0 (a), or L = fa. A mapping f is said to be continuously Frechet di erentiable in U if, at every point x 2 U , the mapping f is Frechet di erentiable and the function x ! f 0 (x) is a continuous mapping of U endowed with the topology induced by the topology of E into the space L(E F ). Let us list some properties of the Frechet derivative. 1. If a mapping f has Frechet derivative at a point a, then this derivative is unique. 2. If f 2 L(E F ), then f 0 (a) = f at every point a 2 E . 3. If f is a mapping of U E into F , g is a mapping of U into F , and both f and g have Frechet derivatives at a point a 2 U , then the function f + g : x ! f (x) + g(x) has Frechet derivative at the point a, which is equal to the sum of the derivatives, (f + g)0 (a) = f 0 (a) + g0 (a). 4. Let E , F , and G be Banach spaces, let U be an open set in E , and let U1 be an open set in F . Let f be a mapping of U into U1 and g a mapping of U1 into G. If f has Frechet derivative f 0 (a) 2 L(E F ) at a point a 2 U and g has Frechet derivative g0 (b) 2 L(F G) at the point b = f (a), then the composition of the mappings h = g f has Frechet derivative at the point a, which is the composition of the derivatives, i.e.,

h0 (a) = g0 (b) f 0 (a) = g0 (f (a)) f 0 (a):

The latter formula is called the chain rule, or the theorem on a composite function. Properties 2 and 3 are obvious. For the proof of Properties 1 and 4, see Schwartz (1967). If E is a product of Banach spaces E1 and E2 , then a mapping f of an open set U E1 E2 into F is a function of two variables, f (x1 x2 ), x1 2 E1 , x2 2 E2 . As x1 is xed at a point a1 , one can consider the partial mapping fa1 : x2 ! f (a1 x2 )

1.9. Di erentiation in Banach spaces

:::

53

and look for the Frechet derivative at a point a2 . If the Frechet derivative of the mapping fa1 at a point a2 exists, it is called the partial Frechet derivative of the @f mapping f with respect to x2 at the point (a1 a2 ) and denoted by @x2 (a1 a2 ). Notice that a mapping f is continuously Frechet di erentiable in U if and @f @f only if it has partial Frechet derivatives @x1 and @x2 which are continuous in U . Then, the Frechet derivative of the mapping f at a point (a1 a2 ) 2 U is given by the formula

@f @f f 0 (a1 a2 )(X1 X2) = @x (a1 a2 )X1 + @x (a1 a2 )X2

1 2

where X1 2 E1 and X2 2 E2 (see Schwartz (1967)).

1.9.2 Implicit function

Let E , F , and G be Banach spaces, let f be a mapping of E F into G, and let c be a point of the space G. Consider the equation

f (x y) = c:

(1:9:1)

Assume that there exists a partial solution of this equation x = a, y = b. Suppose that, for some neighborhood of the point a in E , the equation f (x y) = c with respect to y has a unique solution in a neighborhood of the point b in F . Then, this equation de nes y as a function g(x) of the variable x. g(x) is called the implicit function determined by the equation (1.9.1), and it is characterized by the following property f (x g(x)) = c: (1:9:2) The following theorems on the existence and di erentiability of an implicit function hold (see, e.g., Schwartz (1967)).

Moreover, suppose that Q = @f (a b) is an invertible mapping of F onto G, i.e., @y Q is a bijection and Q;1 2 L(G F ). Then, there exist open sets A and B in the spaces E and F containing a and b, respectively, such that, for any x 2 A, the equation (1.9.1) with respect to y has a unique solution in B . The function y = g(x) determined by this solution is a continuous mapping of A into B .

Theorem 1.9.1 Let E , F , and G be Banach spaces, let U be an open set in E F , and let (a b) 2 U . Let f be a continuous mapping of U into G and let f (a b) = c. Assume that, for every xed x, the function f has the partial Frechet derivative @f (x y ) 2 L(F G) and x y ! @f (x y ) is a continuous mapping of U into L(F G). @y @y

Theorem 1.9.2 Let E , F , and G be Banach spaces, let U be an open set in E F ,

and let f be a mapping of U into G. Let A and B be open sets in the spaces E and F , respectively, A B U , and let g be a mapping of A into B which meets (1.9.2).

54

Chapter 1. Basic de nitions and auxiliary statements

Further, let the mapping f be Frechet di erentiable at a point (a b), b = g(a), let its partial Frechet derivatives

@f (a b) @x

@f (a b) @y

be linear continuous mappings of E and F into G, and let @f (a b) be invertible. @y Then, if the mapping g is continuous at the point a, then it is Frechet di erentiable at this point and the Frechet derivative is given by the relation

g0 (a) = ; @f (a b) @y

;1

@f (a b) : @x

(1:9:3)

1.9.3 The G^teaux derivative and its connection with the a Frechet derivative

Let f be a mapping of an open set U in a Banach space E into a Banach space F , and let a be a point of U . Assume that, for an arbitrary xed h from E , there exists the derivative of the real-valued function t ! f (a + th) at t = 0. This derivative is denoted by f (a h) and is called the G^teaux di erential, or a variation, or derivative of the function f in direction h at the point a. Thus, we have (1:9:4) f (a h) = d f (a + th) = lim f (a + th) ; f (a) :

dt

t=0

The convergence on the right hand side of (1.9.4) is understood in the norm of the space F . The function h ! f (a h) is homogeneous, that is, f (a ch) = c f (a h), c 2 R, however, it is not always linear. Suppose that U = E , i.e., f is a mapping of E into F and there exists an operator L 2 L(E F ) such that f (a h) = Lh for all h 2 E . The operator L is called the G^teaux derivative of the mapping f at the point a and is denoted by a 0 fG (a). Thus

t6=0 t!0 a+th2U

d f (a + th) = f 0 (a)h 0 h 2 E fG (a) 2 L(E F ): G t=0 dt Theorem 1.9.3 Let f be a mapping of a Banach space E into a Banach space F . Let V be an open set in E . Suppose that, at each point x 2 V , there exists the 0 0 G^teaux derivative fG (x) 2 L(E F ). Suppose that x ! fG (x) is a continuous a mapping of V (in the topology generated by the topology of E ) into the space L(E F ). Then, f is a continuously Frechet di erentiable mapping of V into F 0 and its G^teaux derivative is the Frechet derivative, i.e., fG (x) = f 0 (x), x 2 V . a
For the proof see, e.g., Vainberg (1972), Kolmogorov and Fomin (1975).

1.9. Di erentiation of the norm in the space pm ( )

W

55

1.10 Di erentiation of the norm in the space Wpm( )

Apparently, Mazur (1933) was the rst to establish that the norm in the space Lp ( ) is Frechet di erentiable at any nonzero point as long as p > 1. Some items connected with the G^teaux and Frechet di erentiability of the norm are dealt a with in Vainberg (1972). Let us show that, if is a bounded domain in Rn , then the norm in Wpm ( ), p > 1, is a continuously Frechet di erentiable functional everywhere but zero.

1.10.1 Auxiliary statement

that where

Lemma 1.10.1 For any t > 1, there exists a constant c > 0, depending on t, such jz ; yjt c jz jt sign z ; jyjt sign y z y2R (1:10:1)
8 <

;1 sign x = : 0 1 Proof. Introduce the function

if x < 0 if x = 0 if x > 0: (1:10:2)

de ned on the complement to the set of the points of the line z = y in R2 . Dividing the numerator and denominator in (1.10.2) by jz jt z 6= 0, we obtain sign z ; jyj t z (1:10:3) f (z y) = sign z ; y t sign y z that is, the function f depends only on the sign of z and on the ratio jyj . One can z easily see that lim f (z y) = 1 lim f (z y) = 1 lim f (z y) = 1 (y=jzj)!0 (y=jzj)!1 (y=jzj)!;1 (1:10:4) To investigate the behaviour of the function f in the neighbourhood of the line z = y we use the L'Hospital rule of removing indeterminacy. Let z0 be an arbitrary xed point in R, z0 6= 0. Calculating the partial derivative in y of the numerator and denominator in (1.10.3), we have lim f (z0 y) = 0 z0 2 R z0 6= 0: (1:10:5) y !z
0

jt f (z y) = jjz jt signjz ; yyjt sign yj z;j

Now, taking into account (1.10.2){(1.10.5), we get the inequality (1.10.1) with a constant c dependent on t.

56

Chapter 1. Basic de nitions and auxiliary statements

1.10.2 Theorem on di erentiability

in the space Wpm ( ) de ned by the expression

Theorem 1.10.1 Let be a bounded domain in Rn and let f be the functional

f (u) =
X Z
jk j

jDk ujp dx

1 p

= kukWpm( ) :

(1:10:6)

If p > 1, the functional f is continuously Frechet di erentiable in the complement to the zero element of Wpm ( ) and its Frechet derivative is given by

(u)v = kuk1;p

m Wp ( )

X Z

jk j

jDk ujp;2 Dk uDk v dx:

(1:10:7) (1:10:8)

Proof. 1. Denote

'(w) = kwkp pm ( ) : W m ( ). Let u v be arbitrary xed elements of Wp

Introduce the function

g(t) = '(u + tv) =

X Z

jk j

jDk (u + tv)jp dx

t 2 (;1 1):

(1:10:9)

For almost every x 2 , the function under the sign of integral in (1.10.9) is di erentiable in t, and the absolute value of its derivative is majorized by an integrable function in that is independent of t. Hence, applying the theorem on the di erentiability of a function represented as an integral (e.g., Schwartz (1967)), we obtain

dg (0) = d '(u + tv) = X p Z Dk vjDk ujp;1 sign Dk u dx: t=0 dt dt jk j m

(1:10:10)

p 1 Applying the Holder inequality with indices p and q = p;1 ( p + 1 = 1) to the q right hand side of (1.10.10), we get

d dt '(u + tv) t=0

X
jkj

kDk vkLp( ) kDk ukp;(1 ) : Lp

Hence, for a xed u, the right hand side of (1.10.10) de nes a linear continuous functional of v 2 Wpm ( ). Consequently, for all u 2 Wpm ( ), the functional ' de ned by (1.10.8) is G^teaux di erentiable and its G^teaux derivative is given by a a

'0G (u)v =

X
jkj

Dk vjDk ujp;1 sign Dk u dx:

(1:10:11)

1.10. Di erentiation of the norm in the space pm ( )

W

57

2. Now let us show that the function u ! '0G (u) is a continuous mapping of Wpm ( ) into L(Wpm ( ) R) = (Wpm ( )) . Let fung be a sequence such that

(1:10:12) (1:10:13)

un ! u
p p and q = p;1 , we have

in Wpm ( ):

Taking into account (1.10.11) and applying the Holder inequality with indices

j('0G (un ) ; '0G (u)) vj

where
n=
X
jk j

m n kv kWp ( )

(1:10:14)

p ;1 p k un jp;1 sign Dk un ; jDk ujp;1 sign Dk u p;1 dx p : jD

(1.10.15)

Using the inequality (1.10.1) at every point x 2 for which the function p under the sign of integral in (1.10.15) is well-de ned, we conclude that, for t = p;1 , p;1 Z X k un jp sign Dk un ; jDk ujp sign Dk u dx p c1 jD n jk j m X p;1 c1 ( kn + kn ) p (1.10.16)
jkj

where
kn =

Z Z Z

(jDk unjp ; jDk ujp ) sign Dk un dx

=
kn =

jDk un jp ; jDk ujp dx jDk ujp j sign Dk un ; sign Dk uj dx:

fkn = jDk un jp ; jDk ujp :

(1.10.17) (1.10.18) (1:10:19) (1:10:20)

Introduce the notation The continuity of the norm together with (1.10.13) implies that
n!1

lim

f dx = 0 0 kn

jkj = 0 1 2 : : : m

58

Chapter 1. Basic de nitions and auxiliary statements

0 being an arbitrary measurable subset of the set . Known results (Natanson (1974)) and (1.10.20) yield the functions fkn to have uniformly absolutely continuous integrals, that is, for any " > 0 there exists > 0 such that, for any measurable set 0 with mes 0 < (mes standing for the Lebesgue measure), the following estimate holds:

f dx < " 0 kn
0 0

8n:

Further, let = xjx 2 0 = xjx 2 ;

0

fkn (x) 0 fkn (x) < 0 :

Z

jf j dx = 0 fkn dx ; 0 fkn dx 2" 8n 0 kn + ; i.e., the functions jfkn j have uniformly absolutely continuous integrals. Due to (1.10.13) and (1.10.19), we can extract from the sequence ffkn g1 a n=1 subsequence ffkni g1 such that fkni ! 0 a.e. in , and so fkni ! 0 in measure. i=1
Now, because of the Vitali theorem (see Natanson (1974)), we obtain
i!1

Then

lim

jfkni j dx = 0:

(1:10:21)

We may, in turn, extract from any subsequence ffkns g1 a subsubsequence s=1 ffkni g1 that makes (1.10.21) true, so that i=1
n!1

lim

jfkn j dx = 0: jkj = 0 1 2 : : : m: jkj = 0 1 2 : : : m:

(1:10:22)

This equality together with (1.10.17), (1.10.19) implies that lim =0 n!1 kn Using the Lebesgue theorem, we easily see that the numbers kn determined by the formulas (1.10.18) satisfy the following relation: lim =0 n!1 kn (1:10:23) Now, (1.10.12) is a consequence of (1.10.14), (1.10.16), (1.10.22), (1.10.23). 3. On account of (1.10.12) and Theorem 1.9.3, we deduce that the function ' from (1.10.8) is a continuously Frechet di erentiable mapping of Wpm ( ) into R and that '0G (u) = '0 (u) u 2 Wpm ( ): (1:10:24)

1.11. Di erentiation of eigenvalues

59

Obviously the functional f from (1.10.6) is the composition of the mapping ' : Wpm ( ) ! R+ and the function : R+ ! R+ , y ! (y) = y1=p .

The function is continuously di erentiable in R+ n f0g. Now, by using the rule of the di erentiation of a composite function (see subsec. 1.9.1, Property 4), we conclude that the functional f is continuously Frechet di erentiable in the complement to zero in Wpm ( ), and its Frechet derivative is determined by the expression (1.10.7). The theorem is proved.

1.11 Di erentiation of eigenvalues

\ `The owl was a very respectable old bird, terribly well educated,' the mouse said, `she knew more than the night watchman and almost as much as I : : : She proved to me that the night watchman could not hoot unless he used the horn that hung from his shoulder.' " | H. Ch. Andersen \How to Cook Soup upon a Sausage Pin"

1.11.1 The eigenvalue problem

Suppose that we are given spaces U , V such that

U is a Banach space, V is a Hilbert space, V imbedding V ! U is compact.

U , the

(1:11:1)

Let also X be a Banach space, let G be an open set in X , and let h ! ah , h ! bh be functions such that h ! ah is a continuously Frechet di erentiable mapping of (1.11.2) G into L2 (V V R), h ! bh is a continuously Frechet di erentiable mapping of (1.11.3) G into L2 (U U R). We suppose also that, for all h 2 G, the bilinear form ah is symmetric and coercive and the bilinear form bh is symmetric, i.e.,

ah (u v) = ah (v u) u v2V ah (u u) c(h)kuk2 u 2 V c(h) = const > 0 V bh (u v) = bh (v u) u v 2 U:

(1.11.4) (1.11.5) (1.11.6)

60

Chapter 1. Basic de nitions and auxiliary statements

Consider the following eigenvalue problem: (ui (h) i (h)) 2 V R ui (h) 6= 0 v 2 V: i (h)ah (ui (h) v ) = bh (ui (h) v ) (1.11.7)

in which every eigenvalue appears as many times as its multiplicity and every nonzero eigenvalue is of nite multiplicity. The corresponding countable set of eigenfuntions fui(h)g1 forms a basis in the space V , and these functions can be i=1 chosen so that the condition

Theorem 1.11.1 Let the conditions (1.11.1){(1.11.6) be ful lled. Then, for any h 2 G, there exists a countable set f i (h)g1 of eigenvalues of the problem (1.11.7) i=1

ah(ui (h) uj (h)) = ij

is satis ed.

8i j:

(1:11:8)

the bilinear form ah de nes a scalar product in V (see Remark 1.5.2). Now, (1.11.1), (1.11.3), (1.11.6), and Theorem 1.5.8 imply the problem (1.11.7) to have a countable set of eigenvalues f i (h)g1 in which every eigenvalue appears as many times i=1 as its multiplicity and every nonzero eigenvalue is of nite multiplicity. Introduce the operators A(h), B (h) generated by the bilinear forms ah and

Proof. Let h be an arbitrary element of G. By virtue of (1.11.2), (1.11.4), (1.11.5),

bh

(A(h)u v) = ah (u v) (B (h)u v) = bh (u v)

u v2V u v 2 V:

(1.11.9) (1.11.10)

Now, instead of (1.11.7), we have the following problem (ui (h) i (h)) 2 V R ui (h) 6= 0 A(h);1 B (h)ui (h) = i (h)ui (h): (1.11.11) Consider the bilinear form ah as a scalar product in V . Then, by (1.11.6), (1.11.9), and (1.11.10), we get

ah (A(h);1 B (h)u v) = (B (h)u v) = (B (h)v u) = ah (A(h);1 B (h)v u) = ah (u A(h);1 B (h)v)

u v 2 V: (1.11.12)

Hence, A(h);1 B (h) is a selfadjoint operator in V with respect to the scalar product ah( ). Using (1.11.1) and the fact that ah 2 L2 (V V R), bh 2 L2 (U U R), we deduce (see subsec. 1.5.2) that A(h);1 B (h) is a compact operator in V . Known results (see, e.g., Kantorovich and Akilov (1977)) yield that the set of eigenfunctions fui(h)g1 forms a basis in the space V and the condition (1.11.8) is satis ed i=1 provided an appropriate choice of these functions is made.

1.11. Di erentiation of eigenvalues

61

1.11.2 Di erentiation of an operator-valued function

In the proof of Theorem 1.11.1, we de ned the scalar product in V via the bilinear form ah. However, since h runs through the set G, in what follows we will assume that the space V is endowed with a scalar product ( )V that is independent of h. In this case, the relations (1.11.9), (1.11.10) still de ne operators A(h) B (h) 2 L(V V ), and since the operator A(h) is invertible (see Theorem 1.5.2), the problem (1.11.7) reduces to the problem (1.11.11). Notice that the operator A(h);1 B (h) is no longer selfadjoint in V . Introduce the notation

L(h) = A(h);1 B (h):

(1:11:13)

Lemma 1.11.1 Let the conditions (1.11.1){(1.11.6) be satis ed and let operators A(h), B (h), L(h) be determined by the relations (1.11.9), (1.11.10), (1.11.13). Then, the function h ! L(h) is a continuously Frechet di erentiable mapping of G into L(V V ), and at an arbitrary point h0 2 G the Frechet derivative L0 (h0 ) of the function h ! L(h) is given by L0 (h0 )q = A(h0 );1 B 0 (h0 )q ; A(h0 );1 A0 (h0 )q A(h0 );1 B (h0 ) q 2 X:
Here, A0 (h0 ), B 0 (h0 ) are the Frechet derivatives of the functions h ! A(h), h ! B (h) at the point h0 .

(1.11.14)

Proof. Using (1.11.2) and (1.11.9), we easily see that the function h ! A(h) is a continuously Frechet di erentiable mapping of G into L(V V ),
and ((A0 (h0 )q)u v) = a0h0 q(u v)

(1:11:15) (1:11:16)

h0 2 G q 2 X u v 2 V:

Here, a0h0 is the Frechet derivative of the function h ! ah at the point h0 and a0h0 q 2 L2 (V V R). By (1.11.1), (1.11.3), (1.11.10), we obtain that the function h ! B (h) is a continuously Frechet di erentiable mapping of G into L(V V ), and ((B 0 (h0 )q)u v) = b0h0 q(u v) Denote (1:11:17) (1:11:18) (1:11:19)

h0 2 G q 2 X u v 2 V: h2G

f (h) = A(h);1

62

Chapter 1. Basic de nitions and auxiliary statements

and let K (K ;1 ) be the set of the invertible elements of the space L(V V ) (respectively, L(V V )). The function f is the composition of the mapping f1 : h ! f1 (h) = A(h) acting from G into K and the mapping f2 : A ! f2 (A) = A;1 acting from K to K ;1. By virtue of (1.11.15), the function f1 is continuously Frechet di erentiable. The function f2 is a continuously Frechet di erentiable mapping of K into K ;1 (see Schwartz (1967)) and f20 (A)T = ;A;1 T A;1 A 2 K T 2 L(V V ): (1:11:20) By using the theorem on the di erentiation of a composite function (e.g., Schwartz (1967)) we conclude that h ! f (h) = A(h);1 is a continuously Frechet di erentiable mapping of G into L(V V ) (more exactly, into K ;1 ) and f 0 (h0 )q = f20 (f1 (h0 )) f10 (h0 )q = ;A(h0 );1 A0 (h0 )q A(h0 );1 q 2 X: (1.11.21) The function h ! L(h) de ned by the relation (1.11.13) is the composition of the mapping ' : h ! fA(h);1 B (h)g acting from G into L(V V ) L(V V ) and the bilinear continuous mapping '2 : L1 L2 ! L1 L2 acting from L(V V ) L(V V ) into L(V V ). Applying the theorem on the di erentiation of a composite function and noticing (1.11.13), (1.11.17), (1.11.19), we obtain that the function h ! L(h) is a continuously Frechet di erentiable mapping G into L(V V ) and L0 (h0 )q = f (h0 ) B 0 (h0 )q + f 0 (h0 )q B (h0 ) h0 2 G: Therefore, making use of (1.11.19) and (1.11.21), we get (1.11.14), concluding the proof.

1.11.3 Eigenspaces and projections

As before, we suppose that the space V is equipped with a scalar product ( )V that is independent of h. Consider the following eigenvalue problem for the operator L(h) de ned by the relations (1.11.9), (1.11.10), (1.11.13): (ui (h) i (h)) 2 V R ui (h) 6= 0 L(h)ui (h) = i (h)ui (h): (1:11:22) Obviously, the problems (1.11.7) and (1.11.22) have the same eigenvalues and eigenfunctions. By virtue of Theorem 1.11.1, the set of the eigenfunctions of the problem (1.11.22) forms a basis in the space V . So, given h 2 G, an arbitrary element u of V may be uniquely represented as

Projections

u=

1 X

i=1

ci (h u)ui (h)

(1:11:23)

1.11. Di erentiation of eigenvalues

63 (1:11:24) (1:11:25)

where

Here, we suppose the eigenfunctions to satisfy the condition (1.11.8). De ne operators Pi (h) 2 L(V V ) i = 1 2 : : : , by

ci (h u) = ah (u ui ):

Pi (h)u = ci (h u)ui (h):

The operator Pi (h) is the projection onto the one-dimensional subspace generated by the eigenfunction ui (h). However, since the scalar product in the space V is not de ned by the bilinear form ah , Pi (h) is not an orthogonal projection. The formulas (1.11.8), (1.11.23){(1.11.25) yield that the operators Pi (h) to satisfy the following relations:

Pi (h) Pj (h) = ij Pj (h) I being the identity operator in V .

1 X

i=1

Pi (h) = I

(1:11:26)

We will refer to the complexi cation of the space V as V (see, e.g., Kantorovich and Akilov (1977)). An arbitrary element w 2 V is represented in the form w = u + iv, where u v 2 V and i is the imaginary unit. If A 2 L(V V ), setting Aw = A(u + iv) = Au + iAv u v2V (1:11:27) we obtain the operator A 2 L(V V ) which is the complex extension of the operator A. By analogy with (1.11.27), we get the complex extensions of the operators Pi (h), L(h), I , still denoted by the same letters. So,

Pi (h), L(h), I belong to L(V V ) and L(V V ). L(h) ; I =

1 X

(1:11:28) (1:11:29)

Let 2 C . Theorem 1.11.1 and the relations (1.11.23), (1.11.25), (1.11.26) imply
i=1

( i (h) ; )Pi (h):

Thus, if belongs to the complement of the spectrum of the operator L(h) in C , then there exists the operator (L(h) ; I );1 , the inverse of L(h) ; I , and the following formula holds true: (L(h) ; I );1 =
1 X

i=1

( i (h) ; );1 Pi (h):

(1:11:30)

This relation is the consequence of (1.11.26) and (1.11.29). Let (a b) be an interval in R such that i (h) 2 (a b) for i = k k +1 : : : k + l, i (h) < a for i < k , and i (h) > b for i > k + l. Let also ; be the positive oriented

64

Chapter 1. Basic de nitions and auxiliary statements

circle in the complex plane with radius b;a , centered at the point a+b . Then, the 2 2 following relation holds
k +l X i=k

Pi (h) = ; 21 i (L(h) ; I );1 d

;

(1:11:31)

which is easily veri ed by using (1.11.29) and the theorem on the residues of a meromorphic function (e.g., Lavrentiev and Shabat (1973), Schwartz (1967)). Lemma 1.11.2 Let the condition (1.11.1){(1.11.6) be satis ed and let an operator L(h) be determined by the relations (1.11.9), (1.11.10), (1.11.13). Assume that h0 2 G and that is a nonzero eigenvalue of the operator L(h0 ) of multiplicity m 1, i.e., for i 2 J for i 62 J (1.11.32) i (h0 ) = i (h0 ) 6= J = j j +1 ::: j +m ;1 (1.11.33) j being an index. Let also ;0 be a positive oriented circle in the complex plane centered at the point and of so small radius that all the eigenvalues of the operator L(h0) which are not equal to lie outside of ;0 . Then, there exists an open neighbourhood ! of the point h0 in G such that, for any h 2 !, the interior of ;0 contains exactly m eigenvalues (with regard for their multiplicity ) of the operator L(h), and the other eigenvalues of the operator L(h) lie outside of ;0 , i.e., if 0 is the open circle in C with the boundary ;0 , then i2J h2! i (h) 2 0 i 62 J h 2 !: i (h) 2 C n ( 0 ;0 ) The proof of Lemma 1.11.2 is almost identical to the proof of Theorem 1.5.9, so we omit it. Now de ne a function P : ! ! L(V V ) by

P (h) =

;0

(L(h) ; I );1 d :
P

(1:11:34)

The above argument implies that P (h) = i2J Pi (h) and P (h0 ) is the projection onto the eigenspace of the operator L(h0 ) corresponding to the eigenvalue .

Di erentiability of the function P Lemma 1.11.3 Let the conditions (1.11.1){(1.11.6) be ful lled and let operators A(h), B (h), L(h) be de ned by the relations (1.11.9), (1.11.10), (1.11.13). Then, the function h ! P (h) determined by (1.11.34) is a continuously Frechet di erentiable mapping of ! into L(V V ) and of ! into L(V V ).

1.11. Di erentiation of eigenvalues

65

(1:11:35) By virtue of the theorem on the di erentiation of a composite function, we establish, just as in the proof of Lemma 1.11.1 (see (1.11.20), (1.11.21)), that for each 2 ;0 the partial function h ! R(h ) is a continuously di erentiable mapping of ! into L(V V ), and at a point y 2 ! its Frechet derivative @R (y ) is given by @h Here, L0 (y) stands for the Frechet derivative of the function h ! L(h) at a point y, L0 (y) 2 L(X L(V V )). The function h ! R(h ) is the composition of the continuous mapping f1 : h ! f1 (h ) = L(h) ; I acting from ! ;0 into L(V V ) and the continuous mapping f2 : A ! f2(A) = A;1 acting in K1 , the set of the invertible elements of the space L(V V ). Hence, the function h ! R(h ) is a continuous mapping of ! ;0 into L(V V ). Thus, by Lemma 1.11.1 and (1.11.36), we get that
)

Proof. Denote

R(h ) = (L(h) ; I );1 :

@R (y )q = ;R(y ) L0(y)q R(y ) @h

q 2 X:

(1:11:36)

the function h ! @R (h ) is a continuous mapping of (1:11:37) @h ! ;0 into L(X L(V V )). So, if y is an arbitrary element of !, there exists a neighborhood !0 of the point y such that

!0 !

@R (h ) const @h L(X L(V V ))

h 2 !0

2 ;0 :

Now, applying the theorem on the di erentiability of a function represented as an integral (e.g., Schwartz (1967)) to the expression (1.11.34), we conclude that h ! P (h) a continuously Frechet di erentiable mapping of ! into L(V V ). Since P (h) 2 L(V V ), the function h ! P (h) is also a continuously Frechet di erentiable mapping of ! into L(V V ), concluding the proof.

1.11.4 Di erentiation of eigenvalues

G^teaux di erential of the function h ! i (h) a

Let h0 be an arbitrary xed element of G, and let be a nonzero eigenvalue of the operator L(h0 ) of multiplicity m 1, i.e., (1.11.32), (1.11.33) hold. Let also q be an arbitrary xed element of X . There exist numbers a < 0, b > 0 such that h0 + q 2 ! for all 2 (a b), where ! is the open neighbourhood of the point h0 de ned in Lemma 1.11.2. Introduce the notations: ~ L( ) = L(h0 + q) ~i ( ) = i (h0 + q) i 2 J 2 (a b): (1:11:38)

66

Chapter 1. Basic de nitions and auxiliary statements

We suppose that the eigenvalues ~i ( ) i 2 J , are enumerated in such a way that, for arbitrary i i + 1 from J , the following inequalities hold: ~i ( ) ~i ( ) ~i+1 ( ) ~i+1 ( ) for b > for a < 0 0: (1.11.39) (1:11:40)

Then, by (1.11.32) and (1.11.38), we get

Theorem 1.11.2 Let the conditions (1.11.1){(1.11.6) be ful lled and let operators A(h), B (h), L(h) be de ned by the relations (1.11.9), (1.11.10), (1.11.13). Assume that h0 2 G and is an eigenvalue of the operator L(h0 ) of multiplicity m 1, that is, (1.11.32), (1.11.33) hold true. Let also q be an arbitrary element of X and let ~ the eigenvalues of the operator L( ), ~i ( ) i 2 J , (see (1.11.38)) be enumerated so that the inequalities (1.11.39) are valid. Then, the functions ! ~i ( ) are
di erentiable at zero, i.e.,

~i (0) = i (h0 ) =

i 2 J:

~i ( ) = +

(1) + o(

i 2 J: m-dimensional matrix
= lk ]

Moreover, (1) are the eigenvalues of the m i whose elements are de ned by the formula
lk = kl = (((B 0 (h0 ) ;

l k 2 J: (1:11:41) Here, A0 (h0 ), B 0 (h0 ) are the Frechet derivatives of the functions h ! A(h) and h ! B (h) at the point h0 , and ui (h0 ) are eigenfunctions of the operator L(h0) A0 (h0 ))q)ul (h0 ) uk (h0 ))
belonging to the eigenvalue such that

(A(h0 )ul (h0 ) uk (h0 )) = lk In particular, if m = 1, then J = fj g and

(1) = (((B 0 (h

l k 2 J:

(1:11:42) (1:11:43)

Proof. 1. Denote

0) ;

A0 (h0 ))q)uj (h0 ) uj (h0 )):

~ ~ P ( ) = P (h0 + q) Pi ( ) = Pi (h0 + q) ~ ~ M ( ) = P ( )V Mi ( ) =Pi ( )V i 2 J 2 (a b): (1.11.44) Here, the function P is determined by the expression (1.11.34), Pi (h0 + q) is the projection onto the one-dimensional eigenspace of the operator L(h0 + q) belonging to the eigenvalue ~i ( ) (see (1.11.25)). With the notations accepted, the following equality holds true: X ~ ~ P ( ) = Pi ( ) 2 (a b):
i2J

(1:11:45)

1.11. Di erentiation of eigenvalues

67

2 (a b), there exists an operator U ( ) 2 L(V V ) such that

Known results (see, e.g., Kato (1976)) and Lemma 1.11.3 imply that, for any

~ ~ P ( ) = U ( ) P (0) U ( );1 ~i ( ) = U ( ) Pi (0) U ( );1 ~ P (1.11.46) Moreover, ! U ( ) and ! U ( );1 are continuously di erentiable mappings of (a b) into L(V V ) and ;1 ~ ~ ~ P (0) dU (0) L(0) P (0) = 0 d (1.11.47) ~ (0) L(0) dU (0) P (0) = 0 ~ ~ P i 2 J: d For any 2 (a b), U ( ) is a one-to-one mapping of M (0) into M ( ), and of Mi (0) into Mi ( ) for all i 2 J . This operator is called a transforming function. With regard of the notations (1.11.38) and (1.11.44), we get ~ ~ ~ L( ) Pi ( ) = ~i ( )Pi ( ) i 2 J 2 (a b):

(1:11:48)

Multiplying both hand sides of the equality (1.11.48) by U ( );1 from the left and by U ( ) from the right, and noticing (1.11.46), we obtain ^ ~ ~ L( ) Pi (0) = ~i ( )Pi (0) i 2 J 2 (a b) (1:11:49) (1:11:50) By virtue of (1.11.45), the equality (1.11.49) means that ~i ( ), i 2 J , are the ^ ~ eigenvalues of the operator L( ) in the m-dimensional subspace M (0) = P (0)V . 2. By (1.11.38), (1.11.50), taking into account Lemma 1.11.1 and noticing that ! U ( ) and ! U ( );1 are continuously di erentiable mappings of (a b) into L(V V ), we have ^ ! L( ) is a continuously di erentiable mapping of (a b) (1:11:51) into L(V V ). By known results (e.g., Kato (1976)) and (1.11.51), the eigenvalues ~i ( ), i 2 J , are di erentiable at zero and their derivatives dd~i (0) = (1) are the eigenvalues i of the operator ^ ~ ~ T = P (0) dL (0) P (0) d (1:11:52) where ^ ~ L( ) = U ( );1 L( ) U ( ):

in the subspace M (0). By (1.11.50) and (1.11.47), we deduce that ~ ~ ~ T = P (0) dL (0) P (0): d

68

Chapter 1. Basic de nitions and auxiliary statements

~ The eigenfunctions fui(h0 )g, i 2 J , of the operator L(0) form a basis in the subspace M (0). We choose these functions so that they satisfy (1.11.42). Now, we will be occupied with investigation of the representation of the operator T in the matrix form with respect to the basis fui(h0 )g, i 2 J . For convenience of calculation, we assume that the scalar product in V is de ned by the expression (u v)V = (A(h0 )u v) = ah0 (u v): (1:11:53) From (1.11.38) and Lemma 1.11.1, we deduce that ~ dL (0) = L0 (h )q: (1:11:54) 0 d Taking into account (1.11.33), (1.11.42), (1.11.52){(1.11.54), we obtain that the operator T in the basis fui(h0 )g, i 2 J , is represented as the m m-dimensional matrix lk ] whose elements are given by
lk = ((A(h0 )

L0 (h0 )q)ul (h0 ) uk (h0 ))

l k 2 J:

(1:11:55)

From here, noticing (1.11.11), (1.11.14) and (1.11.40), we get

A(h0 );1 B (h0 ))ul (h0 ) uk (h0 )) = (((B (h0 ) ; A (h0 ))q)ul (h0 ) uk (h0 )) l k 2 J: (1.11.56) Since the bilinear forms ah and bh are symmetric, we easily see that ((A0 (h0 )q)u v) = ((A0 (h0 )q)v u) ((B 0 (h0 )q)u v) = ((B 0 (h0 )q)v u) u v 2 V:
0 0

lk = ((B 0 (h0 )q ; A0 (h0 )q

These equalities together with (1.11.56) yield that lk = kl . If is a simple eigenvalue of the operator L(h0 ), then m = 1, J = fj g, and ~ the operator T from (1.11.52) in the one-dimensional subspace M (0) = P (0)V is just the multiplication by the number equal to the right hand side of the equality (1.11.43). The theorem is proved.

Remark 1.11.1 Since the problems (1.11.7) and (1.11.22) are equivalent, Theorem 1.11.2 states that, if the conditions (1.11.1){(1.11.6) are ful lled, then every eigenvalue i (h) of the problem (1.11.7) is di erentiable in any direction, i.e., if h0 is an arbitrary element of G, then for any i and q 2 X there exists the derivative
which is the G^teaux di erential of the function h ! i (h) at the point h0 (see a subsec. 1.9.3). If the multiplicity of an eigenvalue = i (h0 ) equals m > 1, then (1.11.57) has the following sense: there exists an enumerating of the eigenvalues (see (1.11.38), (1.11.39)) such that (1.11.57) holds.

d d i (h0 + q)

=0

= i (h0 q) = (1) i

(1:11:57)

1.11. Di erentiation of eigenvalues

69
lk of the

Remark 1.11.2 Due to (1.11.16), (1.11.18), (1.11.41), the elements

matrix can be represented as
; lk = b0h0 ;

Remark 1.11.3 Let the conditions of Theorem 1.11.2 be ful lled, let the multiplicity m of an eigenvalue is 2, and let J = f1 2g. (It should be noted that the

a0h0 q(ul (h0 ) uk (h0 ))

l k 2 J:

(1:11:58)

case of a double eigenvalue has been considered by Bratus and Seiranian (1983b) in the study of the problem of maximization of the minimal eigenvalue of a selfadjoint operator.) Then the eigenvalues of the matrix , (1) and (1) , are the roots 1 2 of the following equation det or, equivalently,
2;( 11 ; 21 12 22 ;

=0

2 + ( 11 22 ; 12 ) = 0: The roots of the latter equation are given by the formula 11 + 22 ) (1) 12

11 + 22

11 + 22 )2

11 22 +

2 12

1 2

(1:11:59)

Thus, from (1.11.41) it follows that, in the case when the multiplicity of an eigenvalue is greater than 1, the G^teaux di erential i (h0 q) = (1) is not, a i generally speaking, a linear function of q, so that the G^teaux derivative of the a function h ! i (h), i 2 J , at the point h0 does not exist (see subsec. 1.9.3).

Frechet di erentiability Theorem 1.11.3 Let the conditions (1.11.1){(1.11.6) be satis ed, let h0 2 G, and
j (h0 ) = i (h0 ) 6=

let be a nonzero simple eigenvalue of the problem (1.11.7) for h = h0 , i.e., there exists j such that

i 6= j: (1:11:60) Then, there exist an open neighbourhood ! of the point ho in G such that the function h ! j (h) is a continuously Frechet di erentiable mapping of ! into R and its Frechet derivative at a point h 2 ! is given by 0 (h)q = (((B 0 (h) ; q 2 X: (1:11:61) j (h)A0 (h))q )uj (h) uj (h)) j Here, A0 (h), B 0 (h) are the Frechet derivatives at the point h of the functions h ! A(h), h ! B (h) de ned by the relations (1.11.9), (1.11.10), and the eigenfunction uj (h) satisfy the equality ah (uj (h) uj (h)) = 1: (1:11:62)

70

Chapter 1. Basic de nitions and auxiliary statements

Proof. By virtue of (1.11.60), there exists an open circle 0 in the complex plane C with center at and boundary ;0 of so small radius that i (h0 ) 2 C n ( 0 ;0 ) for all i = j . Lemma 1.11.2 yields the existence of an open neighbourhood ! of 6
the point h0 in G such that

(1.11.3), (1.11.9), and (1.11.10), we infer that

h2! i 6= j h 2 !: (1.11.63) Theorem 1.11.2 implies that the function h ! j (h) considered as a mapping of ! into R is G^teaux di erentiable at any point h 2 ! and its G^teaux derivative a a 0 (h) at a point h is given by jG d 0 j G (h)q = d j (h + q ) =0 = (((B 0 (h) ; j (h)A0 (h))q)uj (h) uj (h)) q 2 X: (1.11.64) The eigenfunction uj (h) meets the condition (1.11.62). From (1.11.2),
i (h) 2 C

j (h) 2 0

n(

;0 )

h ! A0 (h) and h ! B 0 (h) are continuous mappings of ! (1:11:65) into L(X L(V V )). Denote by S the set of the eigenfunctions of index j corresponding to all h from ! satisfying the condition (1.11.62), i.e., S = u j u = uj (h) h 2 ! ah (uj (h) uj (h)) = 1 : (1:11:66) We endow the set S with the topology generated by the topology of V . Let us
show that

Lemma 1.11.3 implies that

h ! uj (h) is a continuous mapping of ! into S . Let t0 2 !, ftn g1 n=1 ! and tn ! t0 in X: P (tn )uj (t0 ) ! P (t0 )uj (t0 ) = uj (t0 ) in V ati (uj (ti ) uj (ti )) = 1

(1:11:67) (1:11:68) (1:11:69)

P (tn ) P (t0 ) being the projections onto the one-dimensional subspaces generated by the eigenfunctions uj (tn ), uj (t0 ). We assume that i = 0 1 2 ::: (1:11:70) that is, uj (ti ) 2 S . By (1.11.69), we deduce that P (tn )uj (t0 ) 6= 0 for n su ciently
large, so

cn P (tn )uj (t0 ) = uj (tn )

(1:11:71)

1.12. The Lagrange principle in smooth extremum problems

71

where cn is a constant. Due to (1.11.2), (1.11.68){(1.11.71), we get limn!1 cn = 1. From this and (1.11.69), (1.11.71), we conclude that uj (tn ) ! uj (t0 ) in V as n ! 1, i.e., (1.11.67) holds. The relations (1.11.1){(1.11.6) and Theorem 1.5.9 imply h ! j (h) is a continuous mapping of ! into R. (1:11:72) Using (1.11.64), (1.11.65), (1.11.67), and (1.11.72), we easily see that h ! 0j G (h) is a continuous mapping of ! into X . (1:11:73) Now, (1.11.64), (1.11.73), and Theorem 1.9.3 yield that h ! j (h) is a continuously Frechet di erentiable mapping of ! into R and its Frechet derivative is determined by the relation (1.11.61).

1.12 The Lagrange principle in smooth extremum problems

Let X , Y be Banach spaces and let U be an open set in X . Suppose we are given functionals gi mapping U into R, i = 0 1 2 : : : m, and a function F mapping U into Y . De ne a set Uad in the following way: Uad = u j u 2 U F (u) = 0 gi (u) 0 i = 1 2 : : : m : (1:12:1) Supposing Uad is not empty, consider the following extremum problem: Find a function u such that ^

u 2 Uad ^

g0(^) = u2U g0 (u): u inf

ad

(1:12:2)

The function u is called a point of minimum, or simply a minimum of the problem ^ (1.12.2). A function u 2 Uad is called a point of local minimum, or simply a local ^ minimum of the problem (1.12.2) if there exists a neighbourhood U0 of the point u in U such that ^ g0 (^) = u2Uinf\U g0 (u): u (1:12:3) We place in correspondence to the problem (1.12.2) the following so-called Lagrange functional:
ad 0

L(u w ) =
u 2 U.
where =(
0 1

m X i=0

i gi (u) + (w

F (u))

(1:12:4)

: : : m ) 2 Rm+1 , w 2 Y are the Lagrange multipliers and

The following theorem states necessary conditions for a function to be a local minimum of the problem (1.12.2).

72

Chapter 1. Basic de nitions and auxiliary statements

Theorem 1.12.1 Let X Y be Banach spaces, let U be an open set in X , let F be a continuous mapping of U into Y , and let g0 g1 : : : gm be continuous mappings of U into R. Suppose that a nonempty set Uad is de ned by the relation (1.12.1) and that a function u from Uad is a local minimum of the problem (1.12.2), i.e., ^ u meets the condition (1.12.3), where U0 is a neighbourhood of the point u in U . ^ ^ Assume also that the functionals g0 g1 : : : gm are Frechet di erentiable in U0 , and the mapping F is continuously Frechet di erentiable in U0 . At last, suppose that RF 0 (^) is a closed subspace of Y , RF 0 (^) being the range of the mapping F 0 (^). u u u ^ = (^0 ^1 : : : ^m ) 2 Rm+1 , w 2 Y not Then, there exist Lagrange multipliers ^ all equal to zero and such that
0 : : : ^m 0 m @ L (^ w ^) = X ^ g0 (^) + (F 0 (^)) w = 0 u ^ u ^ i iu @u i=0 ^i gi (^) = 0 u i = 1 2 ::: m ^0 (1.12.5) (1.12.6) (1.12.7)

(F 0 (^)) being the adjoint operator of F 0 (^). Moreover, if RF 0 (u) = Y and there u u exists an element v 2 X such that F 0 (^)v = 0, gi0 (^)v < 0, i = 1 2 : : : m, then u u ^0 6= 0 and without loss of generality one may take ^0 = 1. For the proof, see, e.g., Io e and Tikhomirov (1979). We adduce also the following theorem, see Pshenichny (1980).

Theorem 1.12.2 Let X be a Banach space and let M be a convex set in X . Let
also U be an open set in X and M

U . Assume that u is a solution of the problem ^ g0 (^) = u2U g0 (u) u inf

ad

u 2 Uad ^
where

(1.12.8)

Uad = u j u 2 M gi (u) 0 i = 1 : : : r gi (u) = 0 i = r +1 r +2 : : : r + m

and gi 's are given functionals that are continuously Frechet di erentiable in U . Then, there exist constants i not all equal to zero such that
r+m X i=0
0 u ^ i gi (^)(u ; u)

u2M

(1.12.9) (1.12.10) (1.12.11)

0 i = 0 1 ::: r u i = 1 : : : r: i gi (^) = 0
i

1.13. -convergence and -closedness of linear operators

G G

73

1.13 G-convergence and G-closedness of linear operators

Let V be a separable Hilbert space over R. A linear continuous operator A : V ! V is called coercive if there exists a constant c > 0 such that (Au u) ckuk2 V

u 2 V:

(1:13:1)

Denote by Lc (V V ) the set of linear, continuous, coercive operators acting from V into V . For an operator A 2 Lc (V V ), de ne a bilinear form a on V V in the following way: a(u v) = (Au v) u v 2 V: (1:13:2)

A;1 2 Lc (V V ).

(1.13.1), (1.13.2), and Theorem 1.5.2 imply the existence of the operator

A sequence fAn g Lc (V V ) is said to G-converge to an operator A 2 G Lc (V V ) (notation: An ! A) if lim g A;1 f = g A;1f n n!1
by
; ;

f g2V :

(1:13:3)

Theorem 1.13.1 Let V be a Hilbert space over R and let a set Q(c1 c2) be de ned
Q(c1 c2 ) = A j A 2 L(V V ) (Au v) = (Av u) u v 2 V c1 kuk2 (Au u) c2 kuk2 u 2 V V V
(1.13.4)

c1 , c2 being positive constants, Then, from any sequence fAn g Q(c1 c2 ) one can extract a subsequence fAm g which G-converges to an operator A 2 Q(c1 c2 ):
For the proof, see Zhikov et al. (1993). It should be noted that the G-convergence of elliptic operators is not connected with the convergence of their coe cients, and any kind of convergence of coe cients making the solutions converge in the distribution sense is stronger than the G-converges (see Marcellini (1979), Zhikov et al. (1993)). Further, let Q1 (c1 c2 ) be a subset of Q(c1 c2 ). In particular, Q1 (c1 c2 ) may coincide with Q(c1 c2 ). The set Q1 (c1 c2 ) is called to be G-closed if it contains all limit operators in the sense of the G-convergence, i.e., for any sequence fAn g G Q1 (c1 c2 ) the condition An ! A implies A 2 Q1 (c1 c2 ). Theorem 1.13.1 yields the set Q(c1 c2 ) to be G-closed. Problems of the Gclosedness for second-order elliptic operators have been studied by Tartar (1975), Marcellini (1979), Raitum (1989).

74

Chapter 1. Basic de nitions and auxiliary statements

Let Q be an open or closed set in Rn and let f be a function de ned on Q and taking values in Rk . We say that f is of the C m class if it is m times continuously di erentiable in Q. In the case when Q is a closed set, it means that, in the interior of Q, f has all derivatives of order m and these derivatives coincide with some continuous functions in Q. Let also n = k and let f be a bijection of Q onto f (Q). The bijection f is called a C m -di eomorphism if both f and the inverse bijection f ;1 are of the C m class. We will use the following two theorems, see, e.g., Schwartz (1967).

1.14.1 Di eomorphisms and the relations between the derivatives

1.14 Di eomorphisms and invariance of Sobolev spaces with respect to di eomorphisms

Theorem 1.14.1 Let Q be an open set in Rn and let f be an injection of Q into Rn of the C m class. Suppose that, at an arbitrary point x 2 Q, the Frechet derivative f 0 (x) is an invertible element of the space L(Rn Rn ). Then, f is a C m di eomorphism of Q onto f (Q).

Theorem 1.14.2 Let Q and be open sets in Rn and let f be a bijection of Q onto . Suppose that f and the inverse bijection f ;1 are Frechet di erentiable at every point of Q and , respectively. Then, at each point a 2 Q, the Frechet derivative f 0(a) is a bijection of Rn onto Rn , and the inverse bijection (f 0 (a));1
is the Frechet derivative of f ;1 at the point f (a), i.e.,

(f 0 (a));1 = (f ;1 )0 (b)

b = f (a):
;1

(1.14.1) (1.14.2) (1.14.3)

Notice that, passing to the inverse mappings in (1.14.1), we get

f 0 (a) = (f ;1 )0 (f (a))
or
;

a2Q b2 :

f 0 (f ;1 (b)) = (f ;1 )0 (b)

;1

f ;1 by the derivatives of the function f , while by (1.14.2), (1.14.3) the partial derivatives of f may be expressed by those of f ;1. Let us derive the latter relations. Denote points of Q by y = (y1 : : : yn) and points of by x = (x1 : : : xn ). By fi and ti , i = 1 : : : n, we denote the components of the mappings f and f ;1 , see Fig. 1.14.1, i.e., f = (f1 : : : fn) f ;1 = (t1 : : : tn )
(1.14.4)

The formula (1.14.1) allows one to express the partial derivatives of the function

1.14. Di eomorphisms and invariance of Sobolev spaces

2 @t1 @x1 (x) @t 6 @x2 (x) (f ;1 )0 (x) = 6 1.. 6 4 . @tn @x1 (x)

:::

75

@t1 @x2 (x) @t2 @x2 (x)

.. .. . . @tn (x) : : : @x2

f

::: :::

@t1 3 @xn (x) @t2 @xn (x)7 7 .. 7 : . 5 @tn @xn (x)

(1.14.5)

- 1

Figure 1.14.1: Sets Q, By Cramer's rule, we obtain

; 2

and bijections f , f ;1
3

(f ;1 )0 (x)

;1

= z (x) 6 .. .. .. .. 7 : 4 . . . . 5 a1n (x) a2n (x) : : : ann (x)

a11 (x) a21 (x) : : : an1 (x) 6 a12 (x) a22 (x) : : : an2 (x) 7 6 7

(1.14.6)

@t Here, z (x) = det((f ;1 )0 (x)) ;1 and aij (x) is the cofactor of the element @xij (x) of the matrix (f ;1 )0 (x). We have
2 @f1 @f1 @y1 (y ) @y2 (y ) 6 @f2 (y ) @f2 (y ) 6 @y2 f 0 (y) = 6 @y1.. .. 6 4 . . @fn (y ) @fn (y ) @y1 @y2

::: ::: :::

.. .

@f1 3 @yn (y )7 @f2 @yn (y )7 .. 7 : 5 . 7 @fn (y ) @yn

(1.14.7)

Put y = f ;1 (x), then (1.14.3), (1.14.6), and (1.14.7) yield

@fk (f ;1 (x)) = z (x)a (x) ik @yi Particularly, for n = 2, we get

i k = 1 : : : n:

(1.14.8)

;1 @t @t @t @t z (x) = @x1 (x) @x2 (x) ; @x1 (x) @x2 (x) 1 2 2 1 @t2 (x) @t2 (x) a11 (x) = @x a12 (x) = ; @x

(1.14.9)

76

Chapter 1. Basic de nitions and auxiliary statements

and the relations (1.14.8) take the form

@t a21 (x) = ; @x1 (x)

2

@t a22 (x) = @x1 (x)

1

(1.14.10)

The formula (1.14.8) expresses the partial derivatives of the functions fk by those of the functions tj , j k = 1 2 : : : n. If the conditions of Theorem 1.14.1 are ful lled, then by using the formula (1.14.8) and applying sequentially the theorem on a composite function, we can obtain formulas expressing the partial derivatives of fk of order l m by those of tj . For example,

@f1 (f ;1 (x)) = z (x) @t2 (x) @y1 @x2 @f2 (f ;1 (x)) = ;z (x) @t2 (x) @y @x
1 1

@f1 (f ;1 (x)) = ;z (x) @t1 (x) @y2 @x2 @f2 (f ;1 (x)) = z (x) @t1 (x) @y @x
2 1

(1.14.11)

@ 2 fk (f ;1 (x)) = @ @fk (y) @yj @yi @yj @yi y=f ;1 (x) n @ @f X k (f ;1 (x)) @fl (f ;1 (x)) = @yj l=1 @xl @yi n @ X = (z (x)aik (x)) z (x)ajl (x): l=1 @xl

(1.14.12)

Let Q and be two bounded domains in Rn and let f = (f1 : : : fn) be a C m di eomorphism of Q onto (see Fig. 1.14.1). Suppose u is a function of the C m class de ned on and taking values in R. Then, u = u f is an m times continu^ ously di erentiable function in Q. Let us consider the question on computation of the derivatives of the function u. By the chain rule (see subsec. 1.9.1), the Frechet ^ derivative of u at a point y 2 Q is given by ^ u0 (y)X = u0 (f (y)) f 0 (y)X ^ X 2 Rn : (1.14.13) Sequentially applying the chain rule, we can compute the Frechet derivatives of higher orders. Speci cally, the second order Frechet derivative is given by u00 (y)(X Y ) = u00 (f (y))(f 0 (y)X f 0 (y)Y ) + u0(f (y)) f 00 (y)(X Y ) ^ (1.14.14) n

1.14.2 Sequential Frechet derivatives and partial derivatives of a composite function

X Y 2R : The Frechet derivative of order s m of the function u = u f at a point ^ y 2 Q is given by (see, e.g., Schwartz (1967))

1.14. Di eomorphisms and invariance of Sobolev spaces

:::

77

(1.14.17) where e i denotes that the vector ei appears li times in the parentheses. The relations (1.14.13), (1.14.16), and the equality yield

s! k1 ! k2 ! : : : ks ! (1!)k1 (2!)k2 : : : (s!)ks k1 +2k2 + +sks =s u(k1 +k2 + +ks ) (f 0 )k1 (f 00 )k2 (f (s) )ks (1.14.15) where f (p) = f (p) (y) and u(q) = u(q) (f (y)) are the Frechet derivatives of orders p and q at points y and f (y), respectively. Let fei gn=1 be a basis in Rn such that ei is the unit vector directed along the i yi coordinate axis. We have @u ^ ^0 i = 1 2 ::: n (1.14.16) @yi (y) = u (y)ei @ jlj u ^ ^(jlj) jlj = l1 + + ln jlj m l1 : : : @ynn (y ) = u (y )(e 1 : : : e n ) l @y1 u(s) (y) = ^
X

f 0 (y)ei = @f1 (y) : : : @fn (y) @yi @yi

(1.14.18)

(1.14.18), we get

n @ u (y) = X @u (f (y)) @fk (y) ^ (1.14.19) @yi @yi k=1 @xk Putting X = ej , Y = ei in (1.14.14) and taking into account (1.14.17),

n @ 2u (y) = X @ 2 u (f (y)) @fk (y) @fl (y) ^ @yj @yi @yi @yj k l=1 @xk @xl (1.14.20) n X @u @ 2 fk (y): + @x (f (y)) @y @y j i k=1 k Analogously, the partial derivatives of u of order jlj > 2 (under the condition ^ m jlj) may be calculated by applying (1.14.15), (1.14.17), (1.14.18). These may

also be directly computed by sequential di erentiation of (1.14.20), by applying the chaine rule.

Theorem 1.14.3 Let Q and be bounded domains in Rn with Lipschitz continuous boundaries. Let f be a C m -di eomorphism of Q onto , m 1. Suppose u 2 Wpm ( ), with p 2 1 1). Then, u = u f 2 Wpm (Q) and the mapping u ! u f ^
is an isomorphism of Wpm ( ) onto Wpm (Q).

1.14.3 Theorem on the invariance of Sobolev spaces

78

Chapter 1. Basic de nitions and auxiliary statements

Proof. 1. Suppose, rst, that u 2 C m ( ) and let us prove that kukWpm (Q) ckukWpm( ) ^
Z Z Z

(1.14.21)

where u = u f and c is independent of u. Denote by x and y points of and Q. By ^ the rule of change of variables in a multiple integral (see, e.g., Schwartz (1967)), we have
Q

ju(y)jp dy = ^

ju(x)jp j det(f ;1 )0 (x)j dx c1 ju(x)jp dx

c1 = max j det(f ;1 )0 (x)j:

(1.14.22) (1.14.23)

where
x2 Since u 2 C m ( ), we have u 2 C m (Q), and due to results of subsec. 1.14.2, we get ^

@yl1

@ jlj u (y) = X a (y) @ jsju (f (y)) ^ s @xs1 @xsn l n @ynn 1 1 0<jsj jlj jlj = l1 + + ln jsj = s1 + + sn :

(1.14.24)

Here, 0 < jlj m and as are functions continuous in Q and depending on the derivatives of f of orders jlj, and independent of u. The equality (1.14.24) and the triangle inequality for the norm of Lp ( ) give
Z

@ jlj u (y) p dy ^ l1 @ynn l Q @y1

Z X

p
;1

p jsj = as (f (x)) @xs1@ u sn (x) det(f ;1 )0 (x) dx @xn 1 0<jsj jlj 1 Z X @ jsj u (x) p dx p c2 jlj m @xs1 @xsn n 1 0<jsj jlj

1 p

(1.14.25)

where c2 is independent of u. Now, (1.14.22) and (1.14.25) yield (1.14.21). 2. Next, let u be an arbitrary function from Wpm ( ) and let u = u f . Since ^ the boundary of is Lipschitz continuous, there exists a sequence fu g such that

u 2 C m( )
and due to the above proved

u ! u in Wpm ( )
)

(1.14.26) (1.14.27)

ku kWpm (Q) ku kWpm ( ^

8 u = u f: ^

1.14. Di eomorphisms and invariance of Sobolev spaces

:::

79

We have

@xs1

where c1 is the constant from (1.14.23). By (1.14.24), for an arbitrary w 2 D(Q), (;1)jlj
jlj u (y) l1@ w ln (y) dy ^ @y1 @yn Q Z jsj X = as (y) @xs@ u@xsn (f (y))w(y) dy 1 n Q 0<jsj jlj 1

@ jsj u (f (y)) ; @ jsj u (f (y)) p dy @xsn @xs1 @xsn n n 1 1 p c1 ku ; ukWpm( ) 0 jsj m

(1.14.28)

1 jlj m: (1.14.29)

@ jlj w (y) dy = Z X a (y) @ jsju (f (y))w(y) dy (;1) u(y) l1 ^ s @xs1 @xsn l n @y1 @ynn Q 0<jsj jlj Q 1 w 2 D(Q) (1.14.30) Therefore, the partial distribution derivatives of the function u are still determined ^ by (1.14.24) and u 2 Wpm (Q). It follows from (1.14.24), (1.14.26), and (1.14.28) ^
jlj

By (1.14.26) and (1.14.28), we pass to the limit in (1.14.29), which gives

Z

that

in Wpm (Q): (1.14.31) Now, (1.14.26), (1.14.27), and (1.14.31) give (1.14.21), Analogously, we prove that ku f ;1 kWpm ( ) c2 kukWpm(Q) u 2 Wpm (Q)

u !u ^ ^

and so the mapping u ! u f is an isomorphism of Wpm ( ) onto Wpm (Q).

Let, as above, Q and be two bounded domains in Rn with Lipschitz continuous boundaries, and let f be a C m -di eomorphism of Q onto . By Theorem 1.14.3, the function u ! u = u f is an isomorphism of Wpm ( ) onto Wpm (Q), and the ^ partial derivatives of the function u = u f are de ned by the formula (1.14.24) ^ (in particular, for jlj = 1 and jlj = 2, by (1.14.19) and (1.14.20)). In solution of several problems, we will use the change of variables, and so we have to derive formulas that express the function jlj ^ x ! @ u (f ;1 (x))
l @y11 l @ynn

1.14.4 Transformation of derivatives under the change of variables

80

Chapter 1. Basic de nitions and auxiliary statements

by the derivatives of u and f ;1 . For jlj = 1, setting y = f ;1 (x) in (1.14.19) and taking into account (1.14.8), we get
n @ u (f ;1 (x)) = X @u (x)z (x)a (x): ^ ik @yi k=1 @xk

(1.14.32)

For jlj = 2, setting y = f ;1 (x) in (1.14.20) and taking into account (1.14.8), (1.14.12), we obtain
n @ 2 u (f ;1 (x)) = X @ 2 u (x)(z (x))2 a (x)a (x) ^ ik jl @yj @yi k l=1 @xk @xl (1.14.33) n n X @u X @ + @x (x) (z (x)aik (x)) z (x)ajl (x) : k=1 k l=1 @xl

Notice that the formula (1.14.33) and the formulas for the partial derivatives of higher orders may also be obtained in the following way. Let the function

'(x) =

l @y11

@ jlju ^ ;1 lj @y ln (f (x)) @yj n

(1.14.34)

is known. By (1.14.8) and the chain rule, we get

l @y11 n X @' @ jlj+1 u ^ (f ;1 (x)) = @x (x) @fk (f ;1 (x)) l @yj @yjlj +1 @ynn k=1 k n X @' = @x (x)z (x)ajk (x): k=1 k

(1.14.35)

Chapter 2

Optimal control by coe cients in elliptic systems

\ `The time has come,' the Walrus said, `To talk of many things: Of shoes|and ships|and sealing-wax| Of cabbages|and kings| And why the sea is boiling hot| And whether pigs have wings.' " | Lewis Carrol \Through the Looking-Glass" Almost all this chapter is concerned with problems of control by coe cients in elliptic systems de ned on a bounded domain R2 . Nevertheless, the technique developed is easily transferred to the case of Rn where n > 2. One has just to raise either the smoothness index or the degree of integrability of elements of a Sobolev space which contains the set of admissible controls.

2.1 Direct problem

2.1.1 Coercive forms and operators

Let be a bounded Lipschitz domain in R2 , let (x y) 2 , W = s=1 W l2s ( ), where ls 1, and let V be a closed in nite-dimensional subspace of W with the norm of the space W : Pj 2 L(W L2 ( )) j = 1 2 ::: k (2:1:1) 81

82

Chapter 2. Optimal control by coe cients in elliptic systems

i.e., Pj is a linear, continuous mapping of W into L2 ( ): De ne a set Yp = h j h 2 Wp1 ( ) e1 h e2 where

(2:1:2) (2:1:3)

e1 e2 are positive constants.

Set a family of bilinear, continuous forms ah on the space V depend on a parameter h from Yp , by the expression

V , which
(2:1:4)

ah ( u v ) =
Here

ZZ X k

i j =1

aij (h)(Pi u)(Pj v) dx dy:

k X i j =1

aij 2 C ( e1 e2 ]) i j = 1 2 ::: k aij = aji i j = 1 2 ::: k aij (t) i j c

k X 2 i i=1

(2.1.5) (2.1.6)

2 Rk t 2 e1 e2 ] c = const > 0: (2.1.7)

The formulas (2.1.1){(2.1.5) imply the continuity of the form ah in the following sense:

jah (u v)j c1 kukV kvkV

u v 2 V h 2 Yp c1 = const > 0 u v 2 V:

(2:1:8) (2:1:9)

and due to (2.1.6), the form ah is symmetric, i.e.,

ah (u v) = ah (v u)
ZZ X k

Further, we assume the system of the operators fPj gk=1 to be coercive in V , i.e., j
j =1

(Pj u)2 dx dy c2 kuk2 V

u 2 V c2 = const > 0:

(2:1:10)

The relations (2.1.2){(2.1.4), (2.1.7), (2.1.10) imply the coercivity of the form ah in the following sense:

ah (u u) c3 kuk2 V

u 2 V h 2 Yp c3 = const > 0:

(2:1:11)

2.1. Direct problem

83

2.1.2 Boundary value problem

uh 2 V

Consider the problem: For given elements f 2 V and h 2 Yp (V stands for the dual space of V ), nd a function uh such that

By using the Riesz theorem, or the Lax{Milgram one (see Theorems 1.5.1 and 1.5.2) we get the following theorem: Theorem 2.1.1 Let ah be a bilinear, symmetric form on V V de ned by (2.1.1), (2.1.4){(2.1.6), and let the inequalities (2.1.7), (2.1.10) hold. A set Yp is supposed to be de ned by (2.1.2), (2.1.3). Then, for any h 2 Yp , f 2 V , the problem (2.1.12) has a unique solution.

ah (uh v) = (f v) v 2 V: (2:1:12) Here, (f v) denotes the value of the functional f 2 V at the element v 2 V .

Remark 2.1.1 In fact, the solution uh to the problem (2.1.12) depends not only
on h, but also on f . However, in what follows, an element f is supposed to be xed, so the dependence of the solution on f will not be indicated in the notations, i.e., we will write uh instead of u(h f ). We will need the following statement.

Lemma 2.1.1 Let the conditions of Theorem 2.1.1 be satis ed, let fhng1=1 Yp , n and let un = uhn be a solution to the problem (2.1.12) for h = hn . Then, for p = 2, the condition hn ! h0 weakly in W21 ( ) yields un ! u0 weakly in V , where u0 is the solution to the problem (2.1.12) for h = h0 , and for p > 2 the condition hn ! h0 weakly in Wp1 ( ) implies un ! u0 strongly in V . Proof. 1. Let fhng1=1 Y2 and n hn ! h0 weakly in W21 ( ): (2:1:13)
Introducing the notations

an (u v) = ahn (u v) c3 kunk2 V
Hence, for all n, that

n = 0 1 2 :::

(2:1:14)

and taking into account (2.1.11), (2.1.12), we get

an (un un ) = (f un ) kf kV kunkV :

(2:1:15) Due to (2.1.13), (2.1.15) we may extract a subsequence fhm umg1=1 such m

kunkV const:

hm ! h0

strongly in L2 ( ) and a.e. in

(2.1.16)

84

Chapter 2. Optimal control by coe cients in elliptic systems

um ! u0 weakly in V (2.1.17) u0 being an element of V . Further, let us show that u0 is a solution of the problem (2.1.12) for h = h0 , i.e., by using the notations (2.1.14), we have to show that a0 (u0 v) = (f v) v 2 V: (2:1:18)
From (2.1.2), (2.1.16), we obtain (2:1:19) Therefore, by virtue of (2.1.1), (2.1.5), (2.1.16) and of the Lebesgue theorem, we get, for an arbitrary xed v 2 V and for xed i, j , aij (hm )Pj v ! aij (h0 )Pj v strongly in L2 ( ) as m ! 1: (2:1:20) By (2.1.1), (2.1.17), we have Pi um ! Pi u0 weakly in L2 ( ) as m ! 1: (2:1:21) With the notations (2.1.14), we obtain from (2.1.4), (2.1.20), and (2.1.21) that lim v 2 V: (2:1:22) m!1 am (um v ) = a0 (u0 v ) Since am (um v) = (f v) v2V m 1 (2:1:23) the relation (2.1.22) implies (2.1.18). So, supposing (2.1.13) is true, we have established the existence of a subsequence fumg1=1 of the sequence fun = uhn g1 such that (2.1.17) holds with m n=1 u0 = uh0 . Let us prove that the relation (2.1.17) remains valid for the entire sequence fung, i.e., un ! u0 weakly in V: Assume the contrary. Then, there exist a functional g 2 V , a number " > 0, and a subsequence fuk hk g1 such that k=1 j(g uk ) ; (g u0)j " 8k (2.1.24) hk ! h0 strongly in L2 ( ) and a.e. in (2.1.25) uk ! u weakly in V ~ (2.1.26) u being an element from V . Here, just as before, we denote uk = uhk . Letting k ~ tend to in nity, we get, by virtue of (2.1.25), (2.1.26), a0 (~ v) = (f v) u v 2 V: (2:1:27) However, by Theorem 2.1.1, there exists a unique element u0 in V satisfying (2.1.18). Combining this with (2.1.27), we obtain u = u0 , hence the relations ~ (2.1.24) and (2.1.26) contradict each other.

e1 hn e2

n = 0 1 2 :::

2.1. Direct problem

85 (2:1:28) (2:1:29)

2. Let now fhng1 n=1 Yp and hn ! h0 weakly in Wp1 ( ) p > 2: By virtue of the imbedding theorem (Theorem 1.6.2), we get lim kh ; h0 kC ( ) = 0 n!1 n

C ( ) being the space of continuous functions on .

then (2.1.30) will imply that un ! u0 strongly in V (see Theorem 1.5.3). So, let us establish the equality (2.1.31). We have an (un un ) = (f un): (2:1:32) Setting v = un in (2.1.18), we get, because of (2.1.32), a0 (u0 un) = an (un un ): (2:1:33) By (2.1.30), lim a (u u ) = a0 (u0 u0 ): n!1 0 0 n This equality together with (2.1.33) yields lim a (u u ) = a0 (u0 u0 ): (2:1:34) n!1 n n n By using the notation (2.1.14) and the relations (2.1.1), (2.1.4), we conclude

It follows from the above argument that (2.1.28) implies un ! u0 weakly in V: (2:1:30) By the condition of the lemma, (2.1.8), (2.1.9), (2.1.11) hold , i.e., the form a0 = ah0 de nes a scalar product in V and a norm which is equivalent to the original one in this space. Hence, if we prove that (2:1:31) nlim a0 (un un ) = a0 (u0 u0 ) !1

jan (un un) ; a0 (un un )j

where

ZZ X k

n ckun k2 V n = max max i j (x y)2

i j =1

j(Pi un )(Pj un )j dx dy
c = const > 0
(2.1.35) (2:1:36) (2:1:37)

jaij (hn ) ; aij (h0 )j :

By (2.1.5), (2.1.29), (2.1.36), and Theorem 1.3.10 we, get lim = 0: n!1 n

86

Chapter 2. Optimal control by coe cients in elliptic systems

By virtue of (2.1.30), kunkV const for all n, and therefore (2.1.35) and (2.1.37) yield lim ja (u u ) ; a0 (un un )j = 0: (2:1:38) n!1 n n n Combining (2.1.34) and (2.1.38), we obtain (2.1.31), and so the lemma is proved. Remark 2.1.2 Theorem 2.1.1 and Lemma 2.1.1 remain valid without the assumption that the form ah(u v) is symmetric, i.e., when the condition (2.1.6) does not hold. Indeed, in this case, Theorem 2.1.1 is true by virtue of the Lax{Milgram theorem. Evidently, the relations (2.1.11), (2.1.30), (2.1.31) are satis ed without the assumption (2.1.6). By (2.1.11), we have a0 (un ; u0 un ; u0 ) = a0 (un un) ; a0 (u0 un ) ; a0 (un u0) + a0 (u0 u0 ) Therefore, taking into account (2.1.30) and (2.1.31), we get un ! u0 strongly in V.

c3 kun ; u0 k2 : V

2.2 Optimal control problem

2.2.1 Nonregular control
Basic assumptions
Let us introduce a set of admissible controls by ^ Qad = h j h 2 W21 ( ) khkW21 ( ) c h h h (2.2.1) k (h u h ) 0 k = 1 2 : : : l : Here ^ ^ c, h, h are positive numbers such that e1 < h < h < e2 , e1 (2.2.2) and e2 being the numbers from (2.1.2) h u ! k (h u) are lower semicontinuous functionals given9 > on Y2 V (endowed with the topology generated by the= (2.2.3) product of the weak topology of W21 ( ) and of the weak> topology of V ), k = 1 2 : : : l. The assumption (2.2.3) means that, provided fhn g Y2 , hn ! h weakly in W21 ( ) and fung V , un ! u weakly in V , we have lim inf k (hn un) k = 1 2 : : : l: k (h u ) n!1 We assume the set Qad to be nonempty. It should be stressed that the function uh in the expression of k (h uh) from (2.2.1) is a solution to the problem (2.1.12) for given h.

2.2. Optimal control problem

87

Let us introduce a goal functional h ! f1 (h) satisfying the following condition

h ! f1 (h) is a continuous mapping of Y2 (equipped with the topology induced by the W21 ( )-weak topology) into R.

(2:2:4)

Existence theorem

The optimal control problem consists in nding a function h0 such that

h0 2 Qad

f1 (h0 ) = h2Q f1 (h): inf

ad

(2:2:5)

Theorem 2.2.1 Let ah be a bilinear, symmetric form on V V de ned by the relations (2.1.1), (2.1.4){(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold. Also, let a nonempty set Qad be de ned by (2.2.1), (2.2.2), (2.2.3) and let the goal functional satisfy the condition (2.2.4). Then, the problem (2.2.5) has a solution. Proof. Since the set Qad is not empty by the hypothesis, by virtue of Theorem 1.1.1 and the argument below it about the completed real line R , there exists a sequence fhn g such that hn 2 Qad 8n (2.2.6)
lim f (h ) = h2Q f1 (h): inf n!1 1 n ad (2.2.7) By Theorem 2.1.1, to every element hn there corresponds an element un = uhn giving a solution to the problem (2.1.12) for h = hn . By using (2.1.11), (2.1.12),

and the notations (2.1.14), we get

c3 kunk2 V
Hence,

an (un un ) = (f un ) kf kV kunkV :

(2:2:8)

(2:2:9) By the de nition of the set Qad, the sequence fhng is bounded in W21 ( ). This fact and (2.2.9) yield the existence of a subsequence fhm umg1=1 such that m

kunkV const

8n:

hm ! h0 weakly in W21 ( ) (2.2.10) hm ! h0 strongly in L2 ( ) and a.e. in (2.2.11) um ! u0 weakly in V: (2.2.12) Passing to the limit as m ! 1, we get, just as in the proof of Lemma 2.1.1, v2V where a0 = ah0 . So, we have established that u0 = uh0 . a0 (u0 v) = (f v)
(2:2:13)

88

Chapter 2. Optimal control by coe cients in elliptic systems

Let us show that

By (2.2.1), (2.2.6), (2.2.11), we obtain

h0 2 Qad :

(2:2:14) (2:2:15)
)

The theorem is proved.

At last, by virtue of (2.2.1), (2.2.3), (2.2.6), (2.2.10), and (2.2.12), we get 0 lim inf k (hm um ) k = 1 2 ::: l (2:2:17) k (h0 u0 ) m!1 where u0 = uh0 . Now, (2.2.14) follows from (2.2.15), (2.2.16), and (2.2.17). Taking to notice (2.2.4), (2.2.7), and (2.2.10), we get f1 (h0 ) = mlim f1 (hm ) = h2Q f1 (h): inf !1
ad

c being the number from (2.2.1).

and by (2.2.1), (2.2.6), (2.2.10), and Theorem 1.2.4, kh0 kW21 ( ) lim inf khmkW21 ( m!1

^ h h0 h

(2:2:16)

2.2.2 Regular control

Let us study the optimal control problem when the set of admissible controls is \more regular" than in subsec. 2.2.1. Let a set of admissible controls Uad be given by ^ Uad = h j h 2 Wp1 ( ) khkWp1( ) c h h h k (h uh ) 0 k = 1 2 : : : l : (2.2.18) Here, ^ ^ c, h, h are positive numbers such that e1 < h < h < e2 (e1 , (2.2.19) e2 being the positive numbers from (2.1.2)) and p > 2, 9 h u ! k (h u) is a continuous mapping of Yp V (en-> dowed with the topology generated by the product of the= Wp1 ( )-weak topology and the V -strong topology) into R, > (2.2.20) k = 1 2 : : : l. The set Uad is equipped with the topology induced by the topology of Wp1 ( ) on Uad and is supposed to be nonempty. Let us introduce a goal functional h ! f1 (h) satisfying the following condition h ! f1 (h) is a continuous mapping of Yp (endowed with the (2:2:21) topology induced by the Wp1 ( )-weak topology) into R.

Basic assumptions

2.2. Optimal control problem

89

Existence theorem

c being the constant from (2.2.18).

By virtue of (2.2.18), the sequence fhng is bounded in Wp1 ( ). Choose a subsequence fhm g1=1 such that m hm ! h0 weakly in Wp1 ( ) (2.2.25) hm ! h0 strongly in C ( ): (2.2.26) Let um = uhm be a solution to the problem (2.1.12) when h = hm , i.e., um 2 V am (um v) = (f v) v 2 V m = 0 1 2 ::: (2:2:27) where am (u v) = ahm (u v): (2:2:28) By (2.2.25) and Lemma 2.1.1, we have um ! u0 strongly in V: (2:2:29) Moreover, u0 satis es (2.2.27), (2.2.28) when m = 0. By virtue of (2.2.23) and (2.2.26), we get ^ h h0 h: (2:2:30) By (2.2.23), (2.2.25), we obtain c lim inf khmkWp1 ( ) kh0 kWp1 ( ) (2:2:31) m!1 By using (2.2.18), (2.2.20), (2.2.23), (2.2.25), and (2.2.29), we get 0 mlim k (hm um) = k (h0 u0 ) k = 1 2 : : : l: (2:2:32) !1

Theorem 2.2.2 Let ah be a bilinear symmetric form on V V de ned by the relations (2.1.1), (2.1.4)-(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold true. Let a nonempty set Uad be de ned by (2.2.18)-(2.2.20) and let the goal functional satisfy the condition (2.2.21). Then, the problem (2.2.22) has a solution. Proof. Let fhng1=1 be a minimizing sequence, that is, n hn 2 Uad 8n (2.2.23) f1 (hn ) ! h2U f1 (h): inf (2.2.24)
ad

Let us consider the problem of nding a function h0 such that h0 2 Uad f1(h0 ) = h2U f1 (h): inf
ad

(2:2:22)

From (2.2.30){(2.2.32), it follows that h0 2 Uad. At last, the relations (2.2.21), (2.2.24), and (2.2.25) imply f1 (h0 ) = mlim f1 (hm ) = h2U f1(h) inf !1 which gives the theorem.
ad

90

Chapter 2. Optimal control by coe cients in elliptic systems

2.2.3 Regular problem and necessary conditions of optimality

Modi cation of the restrictions
'1 (h) = max (h ; h(x y))
(x y)2

In the space Wp1 ( ), p > 2, de ne the following functionals ^ '2 (h) = max (h(x y) ; h)
(x y)2

(2:2:33)

h and ^ being the numbers from (2.2.18). Then, the expression (2.2.18) may be h
rewritten as

fore, to regularize the control problem we change the conditions

Uad = h j h 2 Wp1 ( ) khkWp1( ) c '1 (h) 0 '2 (h) 0 k (h uh) 0 k = 1 2 : : : l : (2.2.34) The functionals '1 and '2 from (2.2.33) are not Frechet di erentiable. There'1 (h) 0 '2 (h) 0
(2:2:35)

by the following ones

di erentiable and the conditions (2.2.36) approximate the conditions (2.2.35) sufciently well. Let an "-net be chosen on the set , i.e., a set of points r = fxi yi gr=1 is i chosen so that, for any point (x y) 2 , there exists at least one point (xk yk ) 2 r satisfying ((x ; xk )2 + (y ; yk )2 )1=2 ": De ne functionals Qi by

i = 1 2 : : : 2r (2:2:36) r being a natural number. The functionals Qi are chosen so that they are Frechet

Qi (h) 0

Qi (h) = h ; h(xi yi ) i = 1 2 ::: r ^ Qi (h) = h(xi;r yi;r ) ; h i = r + 1 r + 2 : : : 2r: (2.2.37) Since Wp1 ( ), p > 2, is continuously imbedded into the Holder space (see Triebel (1978)), for any xed p > 2, we can nd an "-net such that the conditions
(2.2.36) approximate the conditions (2.2.35) su ciently well. More precisely, for any > 0, there exists " > 0 such that, for the corresponding "-net, the conditions

khkWp1 (
will imply

Qi (h) 0 '1 (h)

i = 1 2 : : : 2r(") '2 (h) :

(2:2:38) (2:2:39)

2.2. Optimal control problem

91

In what follows, we assume the net to be chosen in such a way that the ^ condition (2.2.38) yields h 2 Yp , that is, max(h ; e1 e2 ; h) (see (2.1.2) and (2.2.19)). Since, for p > 2, the imbedding of Wp1 ( ) into C ( ) is continuous and the mappings Qi from (2.2.37) are a ne functionals, these mappings are Frechet di erentiable in Wp1 ( ) and their Frechet derivatives are determined by the following formulas:

Q0i (h)q = ;q(xi yi ) i = 1 2 : : : r q 2 Wp1 ( ) Q0i (h)q = q(xi;r yi;r ) i = r + 1 r + 2 : : : 2r: (2.2.40) Now, let us replace the set Uad from (2.2.34) with the following one ~ Uad = h j h 2 Wp1 ( ) khkWp1 ( ) c Qi (h) 0 i = 1 2 : : : 2r k (h uh ) 0 k = 1 2 : : : l : (2.2.41) Further, let a functional 0 on the set Yp V be de ned such that u) is a continuous mapping of Yp V (endowed= with the topology generated by the product of the Wp1 ( )weak topology and the V -strong one) into R. Consider the problem of nding a function h0 such that ~ h0 2 Uad 0 (h0 uh0 ) = inf 0 (h uh ): ~
h2Uad

h u!

0 (h

(2:2:42)

(2:2:43)

By using the argument from the proof of Theorem 2.2.2, we establish the following fact.

Theorem 2.2.3 Let ah be a bilinear, symmetric form on V V determined by

the relations (2.1.1), (2.1.4)-(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold ~ let a nonempty set Uad be de ned by (2.2.41), (2.2.37), (2.2.19), (2.2.20) and let (2.2.42) hold true. Then, the problem (2.2.43) has a solution.

Reformulation of the problem

Now, by using Theorem 1.11.1, we will establish necessary conditions of optimality in the problem (2.2.43). To this end, we have to reformulate the problem (2.2.43). Let us introduce the following sets:

X = Wp1 ( ) V p>2 G = h j h 2 Wp1 ( ) e1 < h < e2 U =G V e1 and e2 being the positive numbers from (2.1.2).

(2.2.44)

92

Chapter 2. Optimal control by coe cients in elliptic systems

Since the imbedding of Wp1 ( ) into C ( ) is continuous, G is an open set in Wp1 ( ), hence, U is an open set in X . De ne a function F : U ! V through the formula (h u) 2 U (F (h u) v) = ah (u v) ; (f v) v2V (2:2:45) where ah is the bilinear form de ned in subsec. 2.1.1, and f a xed element from V (see (2.1.12)) for which the problem (2.2.43) is being solved. Let us de ne on the set U functionals g0 (h u) = 0(h u) (2.2.46) g1 (h u) = khkWp1( ) ; c (2.2.47) gi (h u) = Qi;1 (h) i = 2 3 : : : 2r + 1 (2.2.48) gi (h u) = i;2r;1(h u) i = 2r + 2 2r + 3 : : : 2r + l + 1: (2.2.49) Further, let

such that

Uad = (h u) j (h u) 2 U F (h u) = 0 gi (h u) 0 i = 1 2 : : : m (m = 2r + l + 1)g: (2.2.50) Now, the problem (2.2.43) reduces to the following one: Find a pair h0 , u0
(h0 u0 ) 2 Uad

g0 (h0 u0 ) =

Notice that the relations (2.1.12), (2.2.45), (2.2.50), (2.2.51) yield u0 = uh0 .

(h u)2Uad

inf

g0 (h u):

(2:2:51)

Auxiliary statements Lemma 2.2.1 Let ah be bilinear form on V V determined by the relations (2.1.1), (2.1.4)-(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold. Assume that the functions t ! aij (t) are continuously di erentiable on e1 e2 ], that is, aij 2 C 1 ( e1 e2]) i j = 1 2 : : : k: (2:2:52)
Then, the function F determined by (2.2.45) is a continuously Frechet di erentiable mapping of U into V and its Frechet derivative is given by where @F (h u) belongs to the space L(Wp1 ( ) V ) and is determined through the @h relation
k @F (h u)q w = Z Z X daij (h) q(P u)(P w) dx dy i j @h i j =1 dt q 2 Wp1 ( ) w 2 V:

F 0 (h u)(q v) = @F (h u)q + F (h v) + f @h

q 2 Wp1 ( ) v 2 V

(2:2:53)

(2.2.54)

2.2. Optimal control problem

93
hu

Proof. Given a pair (h u) 2 U , de ne an operator

relation ( hu q w) =
ZZ X k

2 L(Wp1 ( ) V ) by the

that

daij (h) q(P u)(P w) dx dy (2.2.55) i j i j =1 dt q 2 Wp1 ( ) w 2 V: Since h 2 G and G is an open set in Wp1 ( ) (see (2.2.44)), there exists r > 0 such h+z 2 G z 2 d(r 0)
)

where

Let q 2 d(r 0). Then, (2.1.4) and (2.2.45) obviously imply the inequality

d(r 0) = z j z 2 Wp1 ( ) kz kWp1(

Z Z

r :

kF (h + q u) ; F (h u) ;
c
k X i j =1

aij (h + q) ; aij h ; daij (h)q (Pi u)2 dx dy dt

hu q kV

(2.2.56)
1=2

Let (x y) be an arbitrary point from . By the mean value theorem, we obtain

aij (h(x y) + q(x y)) ; aij (h(x y)) ; daij (h(x y))q(x y) dt daij (h(x y) + q(x y)) ; daij (h(x y)) q(x y) = dt dt daij kqk 2 (0 1): (2.2.57) ! dt C ( ) kq kC ( )
Here !( ) stands for the continuity modulus:

! daij " = sup daij (t0 ) ; daij (t00 ) ">0 (2:2:58) dt dt dt where the supremum is taken over t0 t00 2 e1 e2], jt0 ; t00 j ". By (2.2.52) and
Theorem 1.3.10, we get

! daij kqkC ( dt

!0

as kqkC ( ) ! 0:

From (2.2.56), (2.2.57), and the continuity of the imbedding of Wp1 ( ) into C ( ), it follows that hu is the partial Frechet derivative of the function F with respect to h, i.e., @F (h u) = (2:2:59) hu: @h

94

Chapter 2. Optimal control by coe cients in elliptic systems

Assume that

fhng1=1 G h2G fun g1=1 V n n hn ! h in Wp1 ( ) un ! u in V:

Using (2.2.55), one can easily get the estimate

(2.2.60)

k(

hn un ; hu ) q kV k X ZZ daij

i j =1
ZZ

(hn )Pi un ; daij (h)Pi u dx dy dt dt

1=2

kqkC ( ) : (2.2.61)

By (2.1.1) and (2.2.52) we obtain

daij (h )P u ; daij (h)P u 2 dx dy 1=2 i dt n i n dt ZZ daij (h )(P u ; P u) 2 dx dy 1=2 dt n i n i ZZ daij (h ) ; daij (h) P u 2 dx dy 1=2 + i dt n dt ckun ; ukV + c1 ! daij khn ; hkC ( ) : (2.2.62) dt

Due to (2.2.60), the right hand side of the inequality (2.2.62) tends to zero. So, (2.2.59){(2.2.61) yield It follows from (2.2.45) that, for any xed h 2 G, the partial function u ! F (h u) is a continuous a ne mapping of V into V . Therefore, it is Frechet di erentiable and its Frechet derivative @F (h u) is given by the formula @u @F (h u)v = F (h v) + f v 2 V: (2:2:64) @u Let the conditions (2.2.60) hold true again. Taking into account (2.1.4), (2.2.45), (2.2.52), and (2.2.64), we get

h u ! @F (h u) is a continuous mapping of U into L(Wp1 ( ) V ). @h

(2:2:63)

@F (h u ) ; @F (h u) v @u n n @u V k 1=2 X ZZ c ((aij (hn ) ; aij (h))Pi v)2 dx dy c1 'n kvkV

i j =1

2.2. Optimal control problem

95

where 'n = max !(aij khn ; hkC ( ) ): ij However, due to (2.2.60), 'n ! 0, so

Now, (2.2.55), (2.2.59), (2.2.63){(2.2.65) imply (see subsec. 1.9.1) that the function F is a Frechet continuously di erentiable mapping of U into V and its Frechet derivative is given through the relations (2.2.53), (2.2.54). Lemma 2.2.2 Let the conditions of Lemma 2.2.1 be satis ed and let h u be an arbitrary pair from U . Then, the operator F 0 (h u) maps Wp1 ( ) V onto the whole space V , that is, RF 0 (h u) = V (h u) 2 U: (2:2:66) Proof. Given a pair h u from U , let us show that, for an arbitrary z 2 V , there exists a pair q v such that (q v) 2 Wp1 ( ) V F 0 (h u)(q v) = z: (2:2:67) By (2.2.45) and Theorem 2.1.1, we conclude the existence of a function v satisfying ^ the conditions v2V ^ (F (h v ) w) + (f w) = ah (^ w) = (z w) ^ v w 2 V: (2:2:68) From here and (2.2.53), we deduce that the pair q = 0 v = v is a solution to the ^ problem (2.2.67), which completes the proof.

h u ! @F (h u) is a continuous mapping of U into L(V V ): @u

(2:2:65)

Necessary conditions of optimality

L( h u w ) =
m X i=0

Let us return to the the problem (2.2.50), (2.2.51). Introduce the Lagrange functional connected with it:
i gi (h u) + (F (h u) w)

(2:2:69)

: : : m ) 2 Rm+1 u w 2 V h 2 G (see (2.2.44)). Theorem 2.2.4 Let ah be a bilinear, symmetric form on V V determined by the
where = (
0 1

relations (2.1.1), (2.1.4), (2.1.6), (2.2.52) and let the inequalities (2.1.7), (2.1.10) hold. A nonempty set Uad is de ned by (2.2.50), (2.2.19), (2.2.20), (2.2.37), (2.2.45), (2.2.47){(2.2.49) and a goal functional g0 by (2.2.46), (2.2.42). Then, there exists a pair h0 , u0 solving the problem (2.2.50), (2.2.51). If the functionals h u ! i (h u) are Frechet di erentiable in U , i = 0 1 2 : : : l, then there exist Lagrange multipliers ^ = (^0 ^1 : : : ^m ) 2 Rm+1 and w 2 V which do not vanish ^ simultaneously and satisfy the following conditions : ^i 0 i = 0 1 2 ::: m (2.2.70)

96

Chapter 2. Optimal control by coe cients in elliptic systems

@ L (h u w ^)q = ^ @gi (h u )q i 0 0 @h 0 0 ^ i=0 @h + @F (h0 u0 )q w = 0 ^ q 2 Wp1 ( ) @h m @ L (h u w ^)v = a (w v) + X ^ @gi (h u )v = 0 v 2 V: 0 0 ^ h0 ^ i 0 0 @u i=0 @u

~ ~ If a pair h, u exists such that ~ ~ ~ ~ (h u) 2 Wp1 ( ) V F 0 (h0 u0)(h u) = 0 ~ ~ gi0 (h0 u0 )(h u) < 0 i = 1 2 ::: m
then ^0 6= 0 and we may take ^0 = 1.

^i gi (h0 u0) = 0
m X

i = 1 2 ::: m

(2.2.71) (2.2.72) (2.2.73)

(2.2.74)

Proof. The problem (2.2.50), (2.2.51) being a reformulation of the problem (2.2.41), (2.2.43), the existence of a solution to the problem (2.2.50), (2.2.51) follows from Theorem 2.2.3. Let us verify that, in the present setting, the conditions of Theorem 1.12.1 hold true. The sets X and U are de ned by (2.2.44), and Y = V . Due to Lemma 2.2.1, the function F is a Frechet continuously di erentiable mapping of U into Y . The equalities (2.2.46), (2.2.49) and the assumption of the theorem imply the functionals g0 g2r+2 g2r+3 : : : g2r+l+1 to be Frechet di erentiable in U. Theorem 1.10.1 and (2.2.47) yield g1 to be a Frechet continuously di erentiable functional in U . Further, from (2.2.37) and (2.2.48), we deduce the functionals g2 g3 : : : g2r+1 to be Frechet continuously di erentiable in U, too. At last, by virtue of Lemma 2.2.2, we get
RF 0 (h u) = V = Y

(h u) 2 U:

(2:2:75)

Now, Theorem 1.12.1 implies the existence of Lagrange multipliers ^ = (^0 ^1 : : : ^m ) 2 Rm+1

w2Y =V ^

which do not vanish simultaneously and satisfy the conditions (2.2.70), (2.2.71), and the following one (cf. subsec. 1.9.1)

@ L (h u w ^)q + @ L (h u w ^)v = 0 @h 0 0 ^ @u 0 0 ^ q 2 Wp1 ( ) v 2 V:

(2.2.76)

2.2. Optimal control problem

97

Setting v = 0 in this equality and taking into account (2.2.69), we obtain (2.2.72). Now, let us take in (2.2.76) q = 0. By virtue of (2.2.69) and (2.2.64), we get
m @ L (h u w ^)v = X ^ @gi (h u )v + (F (h v) + f w) = 0 ^ 0 0 ^ i 0 0 0 @u i=0 @u

v 2 V:
(2.2.77)

Since the form ah is symmetric and (2.2.45) holds, we have (F (h0 v) + f w) = ah0 (v w) = ah0 (w v) ^ ^ ^ which, together with (2.2.77), yields (2.2.73). Finally, due to Theorem 1.12.1, if ~ ~ there exists a pair h u satisfying the conditions (2.2.74), then ^0 6= 0, so we can ^0 = 1. The theorem is proved. take

2.2.4 Nonsmooth (discontinuous) control

On the incorrectness of the initial problem
L1 ( ). More precisely, de ne set a Uad as
Let us consider the optimal control problem when the controls belong to the space ^ Uad = fh j h 2 L1 ( ) h h h a.e. in g (2:2:78)

^ h h being the positive numbers introduced above. Still, we assume the conditions

(2.1.1), (2.1.4){(2.1.7), (2.1.10) are satis ed. By virtue of Theorem 1.5.2, for a given xed element f 2 V and for any h 2 Uad, there exists a unique function uh such that uh 2 V ah (uh v) = (f v) v2V (2:2:79) and the inequality kuhkV const h 2 Uad (2:2:80) holds. We suppose the goal functional is of the form

f1 (h) = kuh ; z k2 V z being a xed element from V . h0 2 Uad hn 2 Uad 8n f1(h0 ) = h2U f1 (h): inf
ad

(2:2:81)

The optimal control problem consists in nding a function h0 such that (2:2:82) (2:2:83)

Let fhng1 be a minimizing sequence, i.e. n=1 inf nlim f1 (hn ) = h2Uad f1 (h): !1

98

Chapter 2. Optimal control by coe cients in elliptic systems

Because of (2.2.78), the sequence fhng is bounded in L1 ( ), and (2.2.80) implies the sequence fuhn g to be bounded in V . This is why we can nd a subsequence fhmg such that ~ hm ! h -weakly in L1 ( ) (2.2.84) uhm ! u weakly in V: ~ (2.2.85) The relation (2.2.84) means that lim m!1
ZZ

hm g dx dy =

ZZ

~ hg dx dy

g 2 L1 ( ):

Concerning the -weak convergence, see, e.g., Yosida (1971). However, the relations (2.2.84) and (2.2.85) are insu cient to pass to the limit in the expression of ahm (uhm v), so we cannot state that the function u coincides ~ with uh. (In subsecs. 2.2.1 and 2.2.2, we passed to the limit in the expression ~ ahm (uhm v) using Lemma 2.1.1, which cannot be applicable now.) For an arbitrary function h 2 Uad, denote by Ah the operator generated by the bilinear form ah in the following way: (Ah u v) = ah (u v)

u v 2 V:

(2:2:86)

Provided (2.1.1), (2.1.4){(2.1.7), and (2.1.10) hold, Ah is a linear, continuous, selfadjoint, coercive operator acting from V into V , and there exist positive numbers c1 , c2 such that

c1 kuk2 V

(Ah u u) c2 kuk2 V

u 2 V h 2 Uad :

(2:2:87) (2:2:88)

For h 2 Uad and g 2 V , there exists a unique function u(h g) such that

u(h g) 2 V

Ah u(h g) = g:

By Theorem 1.13.1 there exists a subsequence fAhk g1 of the sequence k=1 fAhm g1=1 such that Ahk G-converges to a linear, continuous, selfadjoint, coercive m operator A as k ! 1, i.e., lim A;1 g q = A;1 g q k!1 hk
; ;

g q2V :

(2:2:89)

The inequalities (2.2.87) hold for the operator A with the same numbers c1 and c2 . However, the operator A is not, in general, generated by a function h from Uad , that is, there does not exist h 2 Uad such that A = Ah , and the set of the operators fAh g h 2 Uad , is not G-closed (see Section 1.13 and the example below). We denote by a the bilinear form generated by the operator A

a(u v) = (Au v)

u v 2 V:

(2:2:90)

2.2. Optimal control problem

99

The function u from (2.2.85) is a solution to the following problem ~ u2V ~ a (~ v) = (f v) u v2V where a 6= ah . ~ It follows from the preceding argument that the problem (2.2.78), (2.2.79), (2.2.81), (2.2.82) is ill-posed, so we need a new setting of it. One of the possible ways is the usage of smoother controls. For example, we may assume that h 2 Wp1 ( ) with p 2. A problem of such kind was considered above, in subsecs. 2.2.1 and 2.2.2. Another way requires an essential change of the problem. The function h is not a control any more. As a set of admissible controls we take either a set of operators or a set of elements of some functional space such that, to every element of this set, there corresponds an operator and the set of such operators is G-closed. Let us investigate the latter way by considering an example.

Optimization problem for a second-order elliptic equation

Let V = W 1 ( ) and let a bilinear form ah be given by 2

ah (u v) =

ZZ

@u @v @v h @x @x + @u @y dx dy @y

u v 2 W 1 ( ): 2

(2:2:91)

We suppose the set Uad and the goal functional f1 to be de ned by (2.2.78), (2.2.81). The set of the operators fAh g h 2 Uad, determined by the relations (2.2.86), (2.2.91) is not G-closed since a sequence of operators fAhk g hk 2 Uad can Gconverge to an operator Ab 2 L(V V ) of the form (see Marino and Spagnolo (1969)): (Ab u v) ZZ = b11 @u @v + b12 @u @v + b21 @u @v + b22 @u @v dx dy (2.2.92)

@x @x

@x @y

@y @x

@y @y

where bij 2 L1( ), i j = 1 2, b12 = b21 6= 0: Now, let us de ne a set Pad by the formula

Pad = b j b = fbij g i j = 1 2 bij 2 L1 ( ) b12 = b21 h

; 2 2 1+ 2
2 X

To each element b 2 Pad , there corresponds an operator

i j =1

^; 2 bij (x y) i j h 1 +
; ;

2 2

1 2

2 R a.e. in

: (2.2.93)

Ab 2 L W 1 ( ) W 1 ( ) 2 2

100

Chapter 2. Optimal control by coe cients in elliptic systems

determined by (2.2.92). ; Given a xed element f 2 W 1 ( ) , there exists a unique function ub such 2 that ub 2 W 1 ( ) Ab ub = f: (2:2:94) 2 Let the goal functional look like

f1 (b) = kub ; z k2

W 1( ) 2

(2:2:95)

z being a given element from W 1 ( ). The optimal control problem consists in 2 nding an element b(0) = fb(0) g such that ij b(0) 2 Pad f1 b(0) = b2P f1 (b): inf ad
(2:2:96)

Theorem 2.2.5 Let an operator Ab 2 L;W 1( ) ;W 1( ) 2 2 Proof. Let fb(n)g be a minimizing sequence: b(n) 2 Pad 8n
lim f n!1 1

be determined by (2.2.92), and let a set Pad and a goal functional be de ned by (2.2.93), (2.2.95). Then, the problem (2.2.96) has a solution.

b(n)

= b2P f1 (b): inf

ad

(2.2.97) (2.2.98)

By virtue of the compactness theorem (see Marcellini (1979)), we can nd a subsequence fAb(m) g1=1 of the sequence fAb(n) g1 such that m n=1

ub(m) ! ub(0)
where

weakly in W 1 ( ) 2

(2:2:99) (2:2:100)

b(0) 2 Pad :
2

By (2.2.99), we get lim inf kub(m) k 1 m!1 W 2( ) account (2.2.95) and (2.2.99), we obtain lim inf f1 b(m) m!1

kub(0) kW 1 ( ) . Therefore, taking into

f1 b(0) :
(2:2:101)

Now, (2.2.98), (2.2.100), (2.2.101) yield the vector function b(0) to satisfy the condition (2.2.96). The theorem is proved.

2.2. Optimal control problem

101

An example of the existence on a G-nonclosed set

Here we expound an example from Cea and Malanowski (1970). Let V = W 1 ( ), let a bilinear form ah be de ned by the relation (2.2.91) 2 and let the set Uad look like ^ Uad = h j h 2 L1( ) h h h a.e. in
ZZ

h dx dy = c :

(2:2:102)

^ We assume that h mes < c < h mes , which implies that the set Uad is not empty. Let a goal functional be of the form

f1(h) =

ZZ

h @uh @x

+ @uh @y

dx dy

(2:2:103)

uh being a solution to the problem (2.2.79) when f is a xed function from ; 1 W 2( ) .

Theorem 2.2.6 Let V = W 1( ), let a bilinear form ah be de ned by the relation 2

ad

(2.2.91), and let a set Uad and a goal functional f1 be determined by (2.2.102), (2.2.103). Then, there exists a function ~ such that h ~ ~ h 2 Uad f1 (h) = h2U f1(h): inf (2:2:104)

Proof. Let us show that 9 h ! f1 (h) is a lower semicontinuous function that maps= Uad (endowed with the topology generated by the -weak topology of L1 ( )) into R
Suppose that

(2:2:105)

hn 2 Uad

hn ! h0 -weakly in L1 ( ):

(2:2:106)

Taking into account (2.2.102) and (2.2.106), we easily deduce that h0 2 Uad . Introduce the notation un = uhn n = 0 1 2 : : : (2:2:107) By (2.2.79), (2.2.91), and (2.2.107), we get (f u0 ) =
ZZ

hn @un @u0 + @un @u0 dx dy @x @x @y @y

102

Chapter 2. Optimal control by coe cients in elliptic systems

ZZ

h0 @u0 @x

+ @u0 @y

dx dy:

(2.2.108)

In view of (2.2.103) and (2.2.108), we obtain

f1 (hn ) ; f1 (h0 ) =

2 2 hn @un ; @u0 + @un ; @u0 dx dy @x @x @y @y ZZ 2 2 ; (hn ; h0 ) @u0 + @u0 dx dy: (2.2.109) @x @y

ZZ

Now, (2.2.106) and (2.2.109) yield lim inf f1 (hn ) f1 (h0 ): n!1

From (2.2.102) it follows that the sequence fqn g is bounded in L1 ( ). Hence, we can extract a subsequence fqm g such that qm ! q0 -weakly in L1 ( ): (2:2:111) From (2.2.102) and (2.2.111), we get q0 2 Uad. ~ Now, the relations (2.2.105), (2.2.110), (2.2.111) make the function h = q0 satisfy the condition (2.2.104). Remark 2.2.1 The above argument makes it clear that, if the set Uad is sequentially -weakly closed in L1 ( ) and the goal functional is lower semicontinuous, then the optimal control problem has a solution even if the set of the operators corresponding to the set Uad is not G-closed.

Thus, (2.2.105) holds true. Let fqn g be a minimizing sequence: qn 2 Uad lim f (q ) = q2U f1 (q): inf n!1 1 n
ad

(2:2:110)

2.2.5 Some remarks on the use of regular and discontinuous controls

\The natural is rounded, the arti cial is made up of angles : : : Beauty is Nature in perfection circularity is its chief attribute. Behold the full moon, the domes of splendid temples, the huckleberry pie : : : On the other hand, straight lines show that Nature has been de ected." | O. Henry \Squaring the Circle"

2.3. The nite-dimensional problem

103

The results of subsec. 2.2.4 show that, if the state of a system is described by the bilinear form ah from subsec. 2.1.1, then the optimal control problem is ill-posed when one uses controls from L1 ( ), so one has to enlarge the set of bilinear forms to make the set of the operators generated by these forms G-closed. Such enlarging is not always justi ed from the point of view of physics. In fact, a G-limit operator might have no physical meaning. Besides, the class of the admissible functionals determining the set of the admissible controls and the goal functional is much larger in the case of the admissible controls from Wp1 ( ) than that in the case of controls belonging to L1 ( ). We point out that often, in optimization problems of physics and technics, one has to use regular controls. For instance, in problems of control by a function of either the thickness of a plate or a shell, or by the form of the surface of a shell, discontinuous controls cause concentration of stresses, leading to the distruction of the structure. Moreover, mathematical models of plates and shells do not describe the consentration of stresses caused by the discontinuity of the function of thickness and nonsmoothness of the function of the midsurface of the shell. Besides, these models are applicable only in the cases when the function of the thickness of a plate or a shell is smooth and its derivatives are comparatively small. On the other hand, the use of nonsmooth controls when they are considered as admissible ones enables one to expand essentially the set of admissible controls and to get, generally speaking, much bene t in the values of the goal functional in comparison with the case of smooth controls. We note also that the G-convergence is widely used for the averaging of di erential operators see, e.g., Duvaut (1976), Bensoussan et al. (1978), SanchezPalencia (1980), Bakhvalov and Panasenko (1984), Zhikov et al. (1993). In turn, the averaged di erential operators are applied to the optimization of nonhomogeneous media, in particular, composites and structures, see Rozvany (1989), Lurie (1993), Bendsoe (1994), and references therein.

2.3 The nite-dimensional problem

Let us consider the problem of approximation of the solution to the regular control problem formulated in subsec. 2.2.2 by a solution to a nite-dimensional problem. We suppose fHn g1 to be a sequence of nite-dimensional subspaces of n=1 1 ( ) satisfying the limit density condition, i.e., Wp The nite-dimensional problem consists in nding a function hn such that hn 2 Hn \ Uad f1 (hn ) = h2Hinf U f1 (h): (2:3:2) \
n
ad

n!1 h2Hn

lim inf kh ; wkWp1 ( ) = 0

w 2 Wp1 ( ):

(2:3:1)

Theorem 2.3.1 Let the assumptions of Theorem 2.2.2 hold true and let fHng
be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying the condition

104

Chapter 2. Optimal control by coe cients in elliptic systems

(2.3.1). Let also a sequence fqng1 exist such that n=1

qn 2 U ad 8n qn ! h0 in Wp1 ( )

(2.3.3) (2.3.4)

where U ad is the set of interior points of Uad , and h0 a solution to the problem (2.2.22). Then, for each n su ciently large (say, n k), there exists a solution hn to the problem (2.3.2) and

inf nlim f1 (hn ) = f1 (h0 ) = h2Uad f1 (h): !1

(2:3:5)

Moreover, there exists a subsequence fhm g1=1 of the sequence fhng1 k such that m n= hm ! h0 weakly in Wp1 ( ).

Remark 2.3.1 The set Uad de ned by the relations (2.2.18)-(2.2.20) is equipped

with the topology induced by the topology of Wp1 ( ) on Uad. This is why the condition qn 2 U ad is equivalent to the existence of a number "n > 0 such that d("n qn ) Uad, where

Remark 2.3.2 It follows from the proof of Theorem 2.3.1 that the problem (2.3.2) has a solution for any n if the set Hn \ Uad is not empty for any n.
that

d("n qn ) = h j h 2 Wp1 ( ) kh ; qn kWp1 (

"n :

(2:3:6)

Indeed, by virtue of (2.3.3), (2.3.4), we can take h" = qn for n su ciently large. Since h" 2 U ad , there exists a number such that 0< where

Proof of Theorem 2.3.1. For any " > 0, there exists an element h" 2 U ad such kh" ; h0 kWp1 ( ) ": (2:3:7)
" d( h" ) Uad

(2:3:8)

d( h" ) = h j h 2 Wp1 ( ) kh ; h" kWp1 ( ) : (2:3:9) From (2.3.1), we have, for n su ciently large, an element g from Hn exists such that g 2 d( h" ), and so, because of (2.3.8), g 2 Uad . By (2.3.7){(2.3.9), we
get (2:3:10) The above argument implies the existence of a sequence fgn g1 k such that n=
) )

kg ; h0 kWp1 (

kg ; h" kWp1 ( ) + kh" ; h0 kWp1 (

gn 2 Hn \ Uad

2":

(2.3.11)

2.3. The nite-dimensional problem

105 (2.3.12) (2:3:13)

gn ! h0 strongly in Wp1 ( ):
By (2.2.21), (2.3.12), we obtain lim f (g ) = f1 (h0 ): n!1 1 n

Next, the set Hn \ Uad is not empty if n k, since it contains the function gn . Due to (2.2.18), this set is bounded in Wp1 ( ) and in Hn (Hn being endowed with the norm of the space Wp1 ( )). Obviously, the set

Xn = h j h 2 Hn khkWp1 (

^ c h h h

is closed in Hn . By virtue of (2.2.20) and Lemma 2.1.1, the function h ! k (h uh) is a continuous mapping of Yp (endowed with the topology generated by the Wp1 ( )weak one) into R for p > 2. Therefore, the set

Zn = h j h 2 Hn \ Yp k (h uh ) 0 k = 1 2 : : : l
is closed in Hn . Taking into account that Hn \Uad = Xn \Zn , we deduce Hn \Uad to be closed in Hn . Since the space Hn is nite-dimensional and Hn \ Uad is bounded and closed, the set Hn \ Uad is compact. By (2.2.21), h ! f1(h) is a continuous mapping of Hn \ Uad into R. Now, Theorem 1.3.9 implies that the problem (2.3.2) is solvable for any n k. Moreover, if for any n the set Hn \ Uad is not empty, then the problem (2.3.2) is solvable for all n. By (2.2.22), (2.3.2), (2.3.11),

f1 (gn ) f1 (hn ) f1 (h0 ):

From here and (2.3.13), we get
nlim f1 (hn ) = f1 (h0 ) !1

i.e., (2.3.5) holds. By virtue of (2.3.2) and (2.3.5), the elements hn satisfy the relations (2.2.23), (2.2.24). Now, the argument used in the proof of Theorem 2.2.2 shows that we can nd a subsequence fhm g of the sequence fhn g such that hm ! h0 weakly in Wp1 ( ). We note that the problems (2.2.22) and (2.3.2) might have, in general, not unique solution, and from the sequence fhn g of the solutions to the problems (2.3.2) one can choose subsequences fhmg convergent to the solutions of the problem (2.2.22).

106

Chapter 2. Optimal control by coe cients in elliptic systems

2.4 The nite-dimensional problem (another approach)

2.4.1 The set U t
( )

In Section 2.3, in studying the question on the approximation of a solution to the problem (2.2.22) by a solution to the nite-dimensional problem, we supposed the existence of a sequence of interior points of Uad convergent to the solution to the problem (2.2.22). Below, we propose a technique of nding an approximate solution of the problem (2.2.22) that makes no use of this assumption. ; Let t 2 R, 1 t min 2 ; eh1 eh2 , see (2.1.2) and (2.2.19). De ne a set ^

U (t) = h jh 2 Wp1 ( ) khkWp1( ) tc ^ h(2 ; t) h th k (h uh ) t ; 1 k = 1 2 : : : l :

Here, k are the functionals de ned in subsec. 2.2.2 (see (2.2.20)). We equip the set U (t) with the topology induced by the Wp1 ( )-one. Compairing (2.2.18) and (2.4.1), we get

(2.4.1)

Uad U (t)

Uad = U (1):

(2:4:2)

Directly from the proof of Theorem 2.2.2 we get the following fact.

Theorem 2.4.1 Let ah be a bilinear, symmetric form on V V de ned by the relations (2.1.1), (2.1.4){(2.1.6) and let the inequalities ( 2.1.7), (2.1.10) hold let a set U (t) be de ned by the relations (2.4.1), (2.2.19), (2.2.20) and
1 < t min 2 ; e1 e2 : ^ h h (2:4:3)
Also, let the set U (t) be nonempty and let the goal functional satisfy the condition (2.2.21). Then, there exists a function ht such that

ht 2 U (t)

f1 (ht ) = inf(t) f1(h):

h 2U

(2:4:4)

Remark 2.4.1 The right hand side of (2.4.3) is introduced in order to make the inequality e1 h e2 hold for all h 2 U (t) (see (2.1.2)). We need this because the functionals k and f1 are de ned on Yp .
We will refer to the set of the interior points of U (t) as U (t) . Remind that the topology on U (t) is generated by the topology of Wp1 ( ).

Lemma 2.4.1 Let a set U (t) be de ned by the relations (2.4.1), (2.4.3), (2.2.19), (2.2.20), and let h0 be a solution to the problem (2.2.22). Then, h0 2 U (t) .

2.4. The nite-dimensional problem (another approach)

107

Proof. For all h 2 Wp1( ), we obtain khkWp1( ) kh0kWp1 ( ) + kh0 ; hkWp1 ( ) : (2:4:5) By virtue of (2.2.18), (2.2.22), we have kh0kWp1 ( ) c. From here and (2.4.5), we deduce the existence of a constant 1 > 0 such that, for any h 2 d( 1 h0 ), the
inequality is satis ed, d( 1 h0 ) being the ball in Wp1 ( ) centered at h0 , with radius 1 , de ned by (2.3.9). The imbedding of Wp1 ( ) into C ( ) being continuous for p > 2 and h h0 ^ h, there exists a number 2 > 0 such that, for any h 2 d( 2 h0 ), ^ (2 ; t)h h th: (2:4:7) Lemma 2.1.1 and (2.2.20) yield the function h ! k (h uh ) to be a continuous mapping of Yp (equipped with the topology generated by the Wp1 ( )-weak one) into R for p > 2. Hence, noticing that k (h0 uh0 ) 0 k = 1 2 : : : l, we conclude the existence of a number 3 > 0 such that, for any h 2 d( 3 h0 ),
k (h uh )

khkWp1(

tc

(2:4:6)

t;1

k = 1 2 : : : l:

(2:4:8)

Setting = min( 1 consequently, h0 2 U (t) .

2 3 ),

by (2.4.6){(2.4.8) we get that d( h0 )

U (t) ,

Remark 2.4.2 When proving Lemma 2.4.1, we used only the fact that h0 2 Uad. So, under the assumption of Lemma 2.4.1, not only h0 2 U (t) , but also Uad U (t) .

2.4.2 Approximate solution of the problem (2.2.22)

Let fHn g1 be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying n=1 the conditions (2.3.1), and Hn Hn+1 : (2:4:9) Consider the problem of nding a function hn such that

hn 2 Hn \ U (t) f1 (hn ) = inf (t) f1 (h):

h2Hn \U

(2.4.10) (2.4.11)

Theorem 2.4.2 Let the conditions of Theorems 2.2.2 and 2.4.1 be ful lled, let fHn g1=1 be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying the n

conditions (2.3.1), (2.4.9), and let h0 ht be solutions to the problems (2.2.22),

108

Chapter 2. Optimal control by coe cients in elliptic systems

(2.4.4), respectively. Then, for any n su ciently large (say, n (2.4.10), (2.4.11) has a solution hn such that

k), the problem

f1 (ht ) nlim f1 (hn ) f1 (h0 ): (2:4:12) !1 Also, one can choose a subsequence fhmg1=1 from the sequence fhn g1 k such that m n= ~ ~ hm ! h weakly in Wp1 ( ), h 2 U (t), and
~ lim f (h ) = nlim f1 (hn ) = f1 (h): m!1 1 m !1 (2:4:13)

Proof. Due to Lemma 2.4.1, h0 2 U (t) . Consequently, for some > 0, d( h0 ) U (t) , d( h0 ) being de ned by (2.3.9). From here and (2.3.1), we deduce that the set Hn \ U (t) is nonempty for n su ciently large (say, n k) and that a sequence fgng1=k exists such that n gn 2 Hn \ U (t) (2.4.14) 1 ( ): gn ! h0 strongly in Wp (2.4.15)
By virtue of (2.2.21), (2.4.15), we get lim f (g ) = f1 (h0 ): n!1 1 n (2:4:16) The argument similar to the one used in the proof of Theorem 2.3.1 shows that the set Hn \ U (t) is a compact topological space. Hence, by (2.2.21), we conclude the problem (2.4.10), (2.4.11) to be solvable for any n k, and for any n if the set Hn \ U (t) is not empty for all n. The relations (2.4.4), (2.4.11) imply

f1 (ht ) f1 (hn )
Due to (2.4.9), (2.4.11),

8n:

(2:4:17)

Consequently, using (2.4.17), we deduce the sequence of the numbers ff1 (hn )g1 k n= to be convergent and (2:4:18) nlim f1 (hn ) f1 (ht ): !1 By (2.4.11), (2.4.14), we get

f1 (hn+1 ) f1 (hn ):

f1 (gn ) f1 (hn ):

(2:4:19)

Now, (2.4.16), (2.4.18), (2.4.19) yield (2.4.12). By (2.4.10) and (2.4.1), we conclude the sequence fhn g1 k to be bounded in Wp1 ( ). Let us extract from it a n= subsequence fhmg1=1 such that m hm ! ~ weakly in Wp1 ( ) h (2.4.20)

2.4. The nite-dimensional problem (another approach)

109 (2.4.21)

hm ! ~ h

strongly in C ( )

~ Let us show that h 2 U (t) . Lemma 2.1.1 and the relations (2.2.20), (2.4.20) imply ~ ~ lim (h u ) = k (h uh) k = 1 2 : : : l: (2:4:22) m!1 k m m Here, um = uhm is a solution to the problem (2.1.12) when h = hm. By (2.4.1), (2.4.10), (2.4.22), we have ~ ~ k = 1 2 : : : l: (2:4:23) k (h uh ) t ; 1 Combining (2.4.1), (2.4.10), (2.4.21), we obtain ~ ^ h(2 ; t) h th: (2:4:24) Because of (2.4.1), (2.4.10), (2.4.20), we get tc lim inf khm kWp1 ( m!1
)

~ khkWp1 ( ) :

(2:4:25)

~ Now, the relations (2.4.23){(2.4.25) yield h 2 U (t) . At last, (2.2.21) and (2.4.20) imply (2.4.13). Thus, the theorem is proved. We note that the above technique of the construction of approximate solutions to the problem (2.2.22) is also applicable to the problem (2.2.41), (2.2.43).

So far we investigated the approximation of a solution to the optimal control problem by a solution to the nite-dimensional problem when the set of admissible controls is in Wp1 ( ), p > 2. In both approaches considered in Sections 2.3 and 2.4, we used the existence of a sequence of interior points of the set Uad or U (t) converging to the solution h0 of the optimal control problem. In the case when the set of admissible controls Qad is de ned by (2.2.1), i.e., when Qad W21 ( ), and the topology on Qad is generated by the W21 ( )-strong one, the set of the interior points of Qad is empty, so both methods considered above are not applicable. This reason makes us study the general optimal control problem in the case when the set of admissible controls has no interior points. Suppose U is a Banach space and Uad is a closed, bounded set in U (2:4:26) The set Uad is endowed with the topology induced by that of U , and in this topology Uad may contain no interior points. Suppose also that h ! f (h) is a continuous mapping of Uad into R: (2:4:27)

2.4.3 Approximate solution of the optimal control problem when the set U is empty
ad

110

Chapter 2. Optimal control by coe cients in elliptic systems

Theorem 2.4.3 Let the conditions (2.4.26), (2.4.27) be satis ed and let there exist
a function h0 such that

h0 2 Uad

f (h0 ) = h2U f (h): inf

ad

Suppose fHng is a sequence of nite-dimensional subspaces of U such that Hn Hn+1 for all n and the set H1 \ Uad is not empty. Assume there exists a sequence fgng such that

gn 2 Hn \ Uad hn 2 Hn \ Uad
and moreover
n!1

gn ! h0 strongly in U : f (hn ) = h2Hinf f (h) \U

n
ad ad

(2:4:28) (2:4:29) (2:4:30)

Then, for any n, there exists a solution to the nite-dimensional problem

lim f (hn ) = f (h0 ) = h2U f (h): inf

Proof. By the hypothesis of the theorem, the set Hn \ Uad is not empty for each n, and since the space Hn is nite-dimensional, we conclude Hn \ Uad to be a
compact set. From here and (2.4.27), we deduce that the problem (2.4.29) has a solution for any n. By (2.4.27), (2.4.28), we get
nlim f (gn ) = f (h0 ): !1

This equality together with (2.4.29) implies (2.4.30). Apply Theorem 2.4.3 to the problem (2.2.5). Set

U = W21 ( )

Uad = Qad

f (h) = f1 (h)

Qad being de ned by (2.2.1){(2.2.3) and the function f1 satisfying the condition (2.2.4). By virtue of (2.2.1), the set Qad is bounded in W21 ( ). It follows from the proof of Theorem 2.2.1 that the set Qad is sequentially weakly closed, and so it is closed in the topology induced by the W21 ( )-strong one. Let fHn g be a sequence of nite-dimensional subspaces of W21 ( ) satisfying
the conditions
n!1 u2Hn

lim inf ku ; vkW21 ( ) = 0

v 2 W21 ( ) Hn Hn+1 8n

and suppose that the set H1 \ Qad is not empty. Then, Theorem 2.4.3 implies the existence of a function hn for every n such that

hn 2 Qad \ Hn

f1 (hn ) = h2Qinf\H f1 (h)

ad

2.4. The nite-dimensional problem (another approach)

111

and if a sequence fqn g exists such that

qn 2 Qad \ Hn

qn ! h0 strongly in W21 ( )
ad

Remark 2.4.3 All the results established above in Sections 2.1{2.4 remain valid without the assumption the form ah (u v) to be symmetric (see Remark 2.1.2).

h0 meeting the relation (2.2.5), then lim f (h ) = h2Q f1 (h): inf n!1 1 n

2.4.4 On the computation of the functional h ! k (h uh)

Passage to the limit

Solution of the nite-dimensional optimization problems (2.3.2) and (2.4.10), (2.4.11) is connected with the calculation of the functional h ! k (h uh), where h 2 Yp (see (2.1.2)) and the function uh is a solution to the problem ah (uh v) = (f v) v 2 V: (2:4:31) The problem (2.4.31) is in nite-dimensional, since so is the space V . Let, for every h 2 Yp , there be given a sequence fuhmg1=1 of approximate solutions to m the problem (2.4.31) such that uhm ! uh strongly in V as m ! 1: (2:4:32) Then, by (2.2.20), we have lim h 2 Yp : (2:4:33) k (h uh ) = m!1 k (h uhm ) The latter relation can be used when one solves the problems (2.3.2) and (2.4.10), (2.4.11). There exist a number of methods to nd an approximate solution to the problem (2.4.31) providing one with a sequence fuhmg1=1 satisfying the condition m (2.4.32). Below, we will consider the Riesz method.

Reduction of the problem (2.4.31) to a variational one For every h 2 Yp , ah is a bilinear, symmetric, continuous, coercive form on V V ,
and so it determines a scalar product in V and a norm

kvk1 = ah (v v)]1=2
which is equivalent to the original norm of the space V . De ne on the space V a functional h (v ) = ah (v v ) ; 2(f v ):

(2:4:34) (2:4:35)

112

Chapter 2. Optimal control by coe cients in elliptic systems

Let uh be a solution to the problem (2.4.31). For any w 2 V and any 2 R, we have h (uh + w) = h (uh ) + 2 ah (uh w) ; (f w)] + 2 ah (w w): Hence, taking = 1 and using (2.4.31), we obtain w 2 V: h (uh + w) h (uh ) The form ah being coercive, the equality holds only when w = 0. So, if uh is a solution to the problem (2.4.31), the functional (2.4.35) has its minimum in V at v = uh . Conversely, if h (v) takes its minimal value in V on some element uh then h (uh + w) h (uh ) for all w 2 V and all 2 R, that is, for every xed w, the function ! h (uh + w) reaches its minimal value at = 0. But then the derivative of this function vanishes at = 0, i.e., Hence, if the functional (2.4.35) reaches its minimum on an element uh , then uh is a solution to the problem (2.4.31).

d d h (uh + w)

=0

= 2 ah (uh w) ; (f w) = 0

w 2 V:

The Riesz method

De ne a Riesz operator Qh to be the canonical isometry of V onto V de ned, for any f 2 V , by the relation ah (Qh f v) = (f v) v 2 V: (2:4:36) By (2.4.31), (2.4.36), taking into account the existence of a unique solution to the problem (2.4.31), we get uh = Qhf: (2:4:37) Due to (2.4.35){(2.4.37), the functional (2.4.35) may be represented as (2:4:38) h (v ) = ah (v v ) ; 2ah (uh v ) = ah (v ; uh v ; uh ) ; ah (uh uh ): Further, let fVm g1=1 be a sequence of nite-dimensional subspaces of V m satisfying the limit density condition in V , i.e., lim inf ku ; wkV = 0 w 2 V: (2:4:39) m!1 u2V
m

A Riesz approximate solution to the problem (2.4.31) is de ned as a unique element uhm 2 Vm such that inf (2:4:40) h (uhm ) = v2V h (v ):
m

2.4. The nite-dimensional problem (another approach)

113

The element uhm is characterized by the relation

ah(uhm v) = (f v)
From (2.4.38) and (2.4.40),

v 2 Vm :

(2:4:41)

h (uhm ) = ah (uhm ; uh uhm ; uh ) ; ah (uh uh ) = v2V ah (v ; uh v ; uh) ; ah (uh uh): inf m

Hence, the function uhm is an element of the best approximation to the explicit solution uh to the problem (2.4.31) in the space Vm in the norm k k1 determined by the relation (2.4.34). Now, taking into account (2.4.39) and the equivalence of the norms k k1 and the original one of the space V , we establish the relation (2.4.32) to hold true for the Riesz approximate solutions. We formulate the results obtained as

Theorem 2.4.4 Let ah be a bilinear form on V V determined by the relations (2.1.1), (2.1.4){(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold. Let also the set Yp be determined by the relations (2.1.2), (2.1.3) and f 2 V . Assume that fVm g1=1 is a sequence of nite-dimensional subspaces of V m satisfying the limit density condition (2.4.39). For any h 2 Yp and any m, de ne an approximate solution to the problem (2.4.31) as the unique element uhm 2 Vm satisfying (2.4.41). Then
lim m!1 kuhm ; uh kV = 0
where uh is a solution to the problem (2.4.31).

2.4.5 Calculation and use of the Frechet derivative of the functional h ! m (h uh)
Finite-dimensional regular problem and Frechet derivatives
Consider the regular control problem (2.2.41), (2.2.43) supposing the hypothesis of Theorem 2.2.4 to be ful lled. We may use the techniques given in Sections 2.3 and 2.4 in order to get an approximate solution to this problem. By using the technique of Section 2.3, an approximate solution hn to the problem (2.2.41), (2.2.43) is de ned as ~ hn 2 Hn \ Uad inf (2:4:42) 0 (hn uhn ) = 0 (h uh )
~ h2Hn \Uad

Hn being a nite-dimensional subspace of Wp1 ( ). When applying the technique ~ of Section 2.4, we replace the set Uad in (2.4.42) by a bigger one, namely U (t) .

114

Chapter 2. Optimal control by coe cients in elliptic systems

E ective methods of solution of the regular nite-dimensional problem (2.4.42) are the linearization ones (see Cea (1971), Pshenichny (1983), and Haug et al. (1986)). To use them, one needs to calculate the Frechet derivatives of the functionals h ! m (h uh), m = 0 1 2 : : : l, h ! khkWp1( ) , and h ! Qi (h), i = 1 2 : : : 2r (see (2.2.41)). These functionals are thought of as those de ned on the set G determined by (2.2.44). The functional h ! khkWp1 ( ) is Frechet continuously di erentiable in G and its Frechet derivative is given by the formula (1.10.7). The functionals Qi , i = 1 2 : : : 2r, are also Frechet continuously di erentiable and their Frechet derivatives are given by (2.2.40). Now, let us study the question on di erentiability of the functionals h ! m (h uh ).

The function h ! uh

According to the relations (2.1.12), (2.2.45) and Theorem 2.1.1, to every element h 2 G there corresponds a unique function uh being the solution of the following problem: uh 2 V F (h uh ) = 0: (2:4:43) So, a function ' : G ! V is de ned as h ! '(h) = uh: (2:4:44) Theorem 2.4.5 Let ah be a bilinear, symmetric form on V V determined by the relations (2.1.1), (2.1.4), (2.1.6), (2.2.52) and let the inequalities (2.1.7), (2.1.10) hold true. Then the function ' : G ! V de ned by the relations (2.4.44), (2.1.12) (f assumed to be xed in (2.1.12)) is Frechet continuously di erentiable and its Frechet derivative in an arbitrary point q 2 G is given by
; @F ;1 being the inverse mapping of @F (q '(q)), and @F @F being de@u (q '(q )) @u @u @h

'0 (q)h = ; @F (q '(q)) @u

;1

@F (q '(q))h @h

h 2 Wp1 ( )

(2:4:45)

ned by the formulas (2.2.64), (2.2.54). Proof. From (2.4.43) and (2.4.44), it follows that h ! '(h) is the implicit function de ned by the equation F (h u) = 0. To prove the Frechet di erentiability of the function ', we will use Theorem 1.9.2. The function F is thought of as a mapping of the open set U = G V (see (2.2.44)) into V . Let q be an arbitrary element from G. The argument of the proof of Lemma 2.2.1 implies that the mapping F is Frechet di erentiable at the point (q '(q)) 2 U and the partial Frechet derivatives are:

@F (q '(q)) 2 L(V V ) @u

@F 2 L(W 1 ( ) V ): p @h

2.4. The nite-dimensional problem (another approach)

115

The relations (2.2.45), (2.2.64) imply

coercive (see subsec. 2.1.1), from (2.4.46) and the Riesz theorem we conclude that the mapping @F (q '(q)) is invertible and the inverse mapping @u

@F (q '(q))v w = (F (q v) + f w) = a (v w) v w 2 V: (2:4:46) q @u Since for every q from G, the bilinear form aq is symmetric, continuous, and @F (q '(q)) @u
;1

2 L(V V )
;1

is given through the relation

@F (q '(q)) @u
where .

g=u w 2 V:

(2:4:47) (2:4:48)

u2V

aq (u w) = (g w)

Due to Lemma 2.1.1, h ! '(h) = uh is a continuous mapping of G into V . Now, applying Theorem 1.9.2, we deduce that the mapping ' is Frechet di erentiable in an arbitrary point q 2 G and its Frechet derivative is given by (2.4.45). It remains to show that

q ! '0 (q) is a continuous mapping of G into L(Wp1 ( ) V ):

Let By virtue of Lemma 2.1.1, we get Introduce the notations

(2:4:49) (2:4:50) (2:4:51)

fqn g1=1 G n

q0 2 G

qn ! q0 in Wp1 ( ):

'(qn ) ! '(q0 ) in V: An = @F (qn '(qn )) @u Bn = @F (qn '(qn )) @h n = 0 1 2 :::

(2:4:52) (2.4.53) (2.4.54) (2:4:55)

By (2.2.63), (2.2.65), and (2.4.50){(2.4.52), we obtain that

An ! A0 in L(V V ) Bn ! B0 in L(Wp1 ( ) V )
have

By using the well-known result (see, e.g., Schwartz (1967)), from (2.4.53) we

A;1 ! A;1 in L(V V ): n 0

116 Now, we get

Chapter 2. Optimal control by coe cients in elliptic systems

A;1 Bn ; A;1 B0 L(Wp1 ( ) V ) n 0 A;1 (Bn ; B0 ) L(Wp1 ( ) V ) + (A;1 ; A;1 ) B0 L(Wp1 ( ) V ) n n 0 ; ;1 ;1 c kBn ; B0 kL(Wp1 ( ) V ) + An ; A0 L(V V )
and by virtue of (2.4.54), (2.4.55), the right hand side of this inequality tends to zero as n ! 1. Hence, (2.4.45), (2.4.50), and (2.4.52) imply (2.4.49). The theorem is proved.

Remark 2.4.4 Combining (2.4.45) and (2.4.46), we easily conclude that

aq ('0 (q)h w) = ; @F (q '(q))h w @h F (q '0 (q)h) + f = ; @F (q '(q))h @h h 2 Wp1 ( ) w 2 V: h 2 Wp1 ( ):
(2:4:56) The relation (2.4.46) allows us to rewrite (2.4.56) as (2:4:57)

The Frechet derivative of the functional h ! m(h uh) De ne functionals m : G ! R in the following way:
m (h) = m (h uh )

m = 0 1 2 : : : l:

(2:4:58)

Theorem 2.4.6 Let ah be a bilinear, symmetric form on V V determined by the relations (2.1.1), (2.1.4), (2.1.6), (2.2.52) and let the inequalities (2.1.7), (2.1.10) hold true. Assume h v ! m(h v) are Frechet continuously di erentiable mappings of G V into R, m = 0 1 2 : : : l. Then, the functionals m de ned by (2.4.58), (2.1.12) (in (2.1.12) f is supposed to be xed ) are Frechet continuously di erentiable on G, and at an arbitrary point q 2 G the Frechet derivative of the
functional m is determined by the relation
0

k @ m (q u )h + ZZ X daij (q)h(P u ) P v(m) dx dy q i q j m (q )h = @h i j =1 dt

(2.4.59)

the function v(m) being the solution to the following problem

v(m) 2 V

aq v(m) w = ; @@um (q uq ) w w 2 V m = 0 1 2 : : : l:

(2.4.60)

2.4. The nite-dimensional problem (another approach)

117

By virtue of Theorem 2.4.5, (1) is a Frechet continuously di erentiable mapm ping of G into G V . By the hypothesis, (2) is a Frechet continuously di erentiable m mapping of G V into R. Hence, the functional m is Frechet continuously di erentiable in G (see Schwartz (1967)), and the Frechet derivative of the mapping m at an arbitrary point q 2 G is given by the relation
m (q )h =
0

Proof. The functional m is the composition of the mappings (1) : G ! G V , m h ! (1) (h) = fh uhg and (2) : G V ! R, h v ! (2) (h v) = m(h v). m m m

@ m (q u )h + @ m (q u ) '0 (q)h q q @h @u 1 ( ) m = 0 1 2 : : : l: h 2 Wp '0 (q)h 2 V

(2.4.61)

Since

the following formula is true

@ m (q u ) 2 V q @u

@ m (q u ) '0 (q)h = @ m (q u ) '0 (q)h : (2:4:62) q q @u @u Further, let the function v(m) be a solution to the problem (2.4.60). We take in (2.4.60) w = '0 (q)h. Then, bearing in mind that the form aq is symmetric and
making use of (2.4.62), we get

@ m (q u ) '0 (q)h = ;a v(m) '0 (q)h q q @u = ;aq '0 (q)h v(m) :

(2.4.63)

Setting in (2.4.56) w = v(m) and taking into account that '(q) = uq (see (2.4.44)), we have

;aq '0 (q)h v(m) = @F (q uq )h v(m) : @h

By (2.4.63), (2.4.64), and (2.2.54), we obtain

(2:4:64)

@ m (q u ) '0 (q)h = @F (q u )h v(m) q @u @h q ZZ X k da ij (m) dx dy = dt (q)h(Pi uq ) Pj v i j =1

(2.4.65)

Now, (2.4.59) follows from the relations (2.4.61) and (2.4.65), which completes the proof of the theorem.

118

Chapter 2. Optimal control by coe cients in elliptic systems

2.5 Spectral problem

2.5.1 Eigenvalue problem
Assume that U is a Banach space, V V ! U is compact.

U (L2 ( )) , and the imbedding

(2:5:1)

U U depending on parameter h running over the set Yp .

Consider the following eigenvalue problem: (ui i ) 2 V R i ah (ui v ) = bh (ui v )

Let there be de ned a family of bilinear, symmetric, continuous forms bh on

ui 6= 0 v 2 V:

(2.5.2)

Thinking of the bilinear form ah as of a scalar product in V (see Remark 1.5.2) and applying Theorem 1.5.8, we get the following Theorem 2.5.1 Let ah be a bilinear, symmetric form on V V determined by the relations (2.1.1), (2.1.4){(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold. Assume bh to be a bilinear, symmetric, continuous form on U U depending on a parameter h 2 Yp and (2.5.1) to be valid. Then, the problem (2.5.2) has a countable set of eigenvalues f i g. Each eigenvalue in the sequence f i g is repeated as many times as its multiplicity is each nonzero eigenvalue has a nite multiplicity and the following estimate holds :

m = a (u u)=1 bh (u u) inf
h

ah (u u)=1

sup bh (u u) = M:

(2:5:3)

If m 6= 0, then m is an eigenvalue, and if M 6= 0, then M is an eigenvalue. Also, lim = 0. i!1 i

Remark 2.5.1 If the hypothesis of Theorem 2.5.1 is valid, the form bh is strictly positive on V , i.e., bh(u u) 0 u2V (2:5:4)

and the equality takes place only for u = 0, then = 0 cannot be an eigenvalue of the problem (2.5.2). In this case, all the eigenvalues of the problem (2.5.2) are positive. They form a sequence converging to zero and every eigenvalue is of nite multiplicity.

2.5.2 On the continuity of the spectrum

Let us denote

N1 = a j a 2 L2 (V V R) a(u v) = a(v u) u v 2 V

2.5. Spectral problem

119 (2.5.5) (2.5.6)

a(u u) c(a)kuk2 u 2 V c(a) = const > 0 : V N2 = b j b 2 L2 (U U R) b(u v) = b(v u) u v 2 U :

The set N1 is endowed with the topology induced by the L2 (V V R) one and in this topology N1 is a metric space. If a1 a2 2 N1 , then d(a1 a2 )|the distance between a1 and a2 |is de ned, because of Remark 1.5.3, by the relation

d(a1 a2 ) = sup ja1 (u u) ; a2 (u u)j

kukV =1

(2:5:7)

N2 is a normed space equipped with the norm kbkN2 = sup jb(u v)j:
kuk

U =1 kvkU =1

(2:5:8)

Lemma 2.5.1 Let ah be a bilinear, symmetric form on V V determined by the relations (2.1.1), (2.1.4){(2.1.6) and let the inequalities (2.1.7), (2.1.10) hold. Then h ! ah is a continuous mapping of Yp (endowed with the topology induced by the Wp1 ( )-weak one ) into N1 when p > 2. Proof. Let fhng1=1 Yp p > 2 and hn ! h0 weakly in Wp1( ). Then, by the n
imbedding theorem (see Theorem 1.6.2)

hn ! h0

strongly in C ( ):

(2:5:9)

By (2.1.1), (2.1.4), and (2.1.5), we have

jahn (u u) ; ah0 (u u)j

= Here, c = const > 0 and

ZZ X k

i j =1 ckuk2 n V

(aij (hn ) ; aij (h0 )) (Pi u)(Pj u)) dx dy

u 2 V:

(2.5.10) (2:5:11)

n = sup sup jaij (hn ) ; aij (h0 )j: i j x2

The relation (2.1.5) and Theorem 1.3.10 imply t ! aij (t) to be a uniformly continuous mapping of the interval e1 e2 ] into R. Hence, by (2.5.9), (2.5.11), we obtain limn!1 n = 0. Now, (2.5.7) and (2.5.10) yield
nlim d (ahn ah0 ) = 0 !1

concluding the proof of the lemma.

120

Chapter 2. Optimal control by coe cients in elliptic systems

The dependence of the bilinear form bh on a parameter h running over the set Yp de nes the mapping h ! bh . We suppose this mapping to satisfy the following

h ! bh is a continuous mapping of Yp (endowed with the (2:5:12) topology generated by the Wp1 ( )-weak one) into N2 . The relation (2.5.12) means that, provided fhng Yp and hn ! h weakly in Wp1 ( ), we get kbhn ; bh kN2 ! 0 as n ! 1.

condition:

Recall that the vector normed space of bounded number sequences converging to zero is referred to as `1 0. The norm of an element = f 1 2 : : : g 2 `1 0 is de ned as k k`1 0 = sup j i j: (2:5:13) i Since the bilinear forms ah and bh depend on a parameter h 2 Yp , the eigenvalues i of the problem (2.5.2) depend on h, too, i.e., i = i (h). By virtue of Theorem 2.5.1, to every h 2 Yp there corresponds an element (h) = f 1 (h) 2 (h) : : : g 2 `1 0 i (h) being the eigenvalues of the problem (2.5.2) ordered so that j 1 (h)j j 2 (h)j : (2:5:14) Under such ordering, if the set of the nonzero eigenvalues is countable (i.e., the number of the nonzero eigenvalues is not nite) and zero is an eigenvalue, then the set f i (h)g1 does not contain zero. i=1 Thus, we de ned the mapping h ! (h) of the set Yp into `1 0 . The following statement is valid. Theorem 2.5.2 Let the assumptions of Theorem 2.5.1 be satis ed and (2.5.12) holds. Then, h ! (h) is a continuous mapping of Yp (endowed with the topology induced by the Wp1 ( )-weak one ) into `1 0 when p > 2. Proof. Let fhng Yp , p > 2, and hn ! h weakly in Wp1( ). Then, Lemma 2.5.1 and (2.5.12) imply that ahn ! ah in N1 bhn ! bh in N2 : Now, Theorem 2.5.2 is a consequence of Theorem 1.5.9 and Remark 1.5.5.

2.6 Optimization of the spectrum

De ne a set of admissible controls as

2.6.1 Formulation of the problem and the existence theorem

Uad = h j h 2 Wp1 ( ) khkWp1 (
)

ZZ

^ hg dx dy C1 h h h : (2:6:1)

2.6. Optimization of the spectrum

121

Here,

p > 2, C , C1 , h, and ^ are positive numbers and e1 h < h ^ h e2 , where e1, e2 are the constants from (2.1.2) g 2 L1 ( ) g c = const > 0 a.e. in :
We suppose that the set Uad is nonempty.

(2.6.2) (2.6.3) (2:6:4) (2:6:5) (2:6:6)

Further, assume that we are given a mapping A such that

! A is a continuous mapping of `1 0 into R :

De ne a goal functional as

f (h) = A (h):

Here, (h) = f 1 (h) 2 (h) : : : g and i (h) = i are the eigenvalues of the problem (2.5.2). The optimal control problem consists in nding a function h0 such that

h0 2 Uad

f (h0 ) = h2U f (h): inf

ad

(2:6:7)

Let us consider a few realizations of the mapping A. I. Let A = k k`1 0 = supi j i j. In this case, the condition (2.6.5) is satis ed and the problem (2.6.7) means the minimization on Uad of the absolute value of the rst eigenvalue of the problem (2.5.2). II. Let i 2 R, i 6= 0, i = 1 2 : : : m. Also, let t ! 'i (t) be continuous mappings of R into R such that 'i (t) 0, 'i ( i ) 6= 0, the support of each 'i belongs to a su ciently small neighborhood of the point i and does not contain R zero, R 'i (t) dt = 1, and the supports of the functions 'i and 'j are disjoint when i 6= j . De ne a mapping A by

2 `1 0

A =

m 1 XX i=1 j =1

ci 'i ( j )

ci = const > 0:

(2:6:8)

Since the support of 'i does not contain zero, the summation in (2.6.8) is taken over a nite number of indices and the mapping A meets the condition (2.6.5). The problem (2.6.7) means that we search the h0 for which the number of the eigenvalues contained in given neighborhoods of the points i , which are de ned by the supports of the functions 'i , is minimized. In other words, the solution to the problem (2.6.7) de nes an element h0 for which the eigenvalues of the problem (2.5.2) are \maximally shifted" from the given numbers i .

122

Chapter 2. Optimal control by coe cients in elliptic systems

Theorem 2.6.1 Let ah be a bilinear form on V V determined by the relations (2.1.1), (2.1.4){(2.1.7), (2.1.10). Also, let a mapping h ! bh satisfying the condition (2.5.12) be de ned. Suppose that a set Uad is determined by the relations (2.6.1){ (2.6.3) and satis es (2.6.4), and a goal functional f (h) is de ned by (2.6.5), (2.6.6). Then, the problem (2.6.7) has a solution. Proof. Let fhng1=1 be a sequence such that n hn 2 Uad
n!1

lim f (hn ) = h2U f (h): inf

ad

(2.6.9) (2.6.10)

By virtue of (2.6.1), (2.6.2), the sequence fhng1 is bounded in Wp1 ( ) when n=1 p > 2. Hence, we may nd a subsequence fhmg1=1 of it such that m

hm ! z hm ! z

weakly in Wp1 ( ) strongly in C ( ): in `1 0 :

(2.6.11) (2.6.12) (2:6:13) (2:6:14)

By (2.6.11) and by Theorem 2.5.2, we get (hm ) ! (z ) Taking into account (2.6.5), (2.6.6), (2.6.10), we obtain
m!1

lim f (hm ) = f (z ) = h2U f (h): inf

ad

By (2.6.9) and (2.6.12), we have ^ h z h: (2:6:15) Using (2.6.3), (2.6.9), (2.6.12), and the Lebesgue theorem (Theorem 1.6.1), we get ZZ ZZ lim hm g dx dy = zg dx dy C1 (2:6:16) m!1 where C1 is the number from (2.6.1). By (2.6.9), (2.6.11), we obtain

C lim inf khmkWp1 ( m!1

kz kWp1 (

(2:6:17)

where C is the number from (2.6.1). The relations (2.6.1), (2.6.15){(2.6.17) imply z 2 Uad. Finally, from (2.6.14) we deduce that the function h0 = z is a solution to the problem (2.6.7).

2.6. Optimization of the spectrum

123

2.6.2 Finite-dimensional approximation of the optimal control problem

Lemma on the approximation from the interior of Uad
We equip the set Uad de ned by the relations (2.6.1){(2.6.3) with the topology generated by the Wp1 ( )-one. We will refer to the set of the interior points of Uad as U ad .

Lemma 2.6.1 Let a set Uad be de ned by the relations (2.6.1){(2.6.3) and meet
n!1

the condition (2.6.4). Also, let h be an arbitrary function from Uad that is not identically equal to h in . Then, there exists a sequence fhng U ad such that

lim khn ; hkWp1 ( ) = 0:

(2:6:18)

the functions from the space Wp1 ( ) are equivalent to continuous functions in when p > 2. By (2.6.1), we get h h, and since h is not identically equal to h by the hypothesis, we infer from (2.6.3) that
ZZ

Proof. 1. By virtue of the imbedding theorem for a two-dimensional domain ,

ZZ ZZ ZZ

hg dx dy >

hg dx dy

h dx dy >

h dx dy:

Then, there exists a constant t such that ^ h>t>h

ZZ

hg dx dy > h dx dy >

ZZ

tg dx dy > t dx dy >

ZZ ZZ

hg dx dy h dx dy:

(2.6.19) (2.6.20) (2.6.21)

ZZ

ZZ

Represent the element h as

h = t + q:
Then, on account of (2.6.20) and (2.6.21), we obtain
ZZ

(2:6:22) (2.6.23) (2.6.24)

qg dx dy > 0 q dx dy > 0:

ZZ

124

Chapter 2. Optimal control by coe cients in elliptic systems

De ne functions hn as follows: hn = t + ln q
ZZ

0 < ln < 1

lim l = 1: n!1 n
ZZ

(2:6:25)

The relations (2.6.22), (2.6.23), and (2.6.25) yield

hg dx dy ;

ZZ

hn g dx dy = (1 ; ln ) hn g dx dy < C1

qg dx dy > 0:
(2:6:26)

Hence,

ZZ

8n

where C1 is the number from (2.6.1). Taking into account that h h (2.6.25) imply

^ h, the relations (2.6.19), (2.6.22), and

Let us suppose that the inequality

^ h < hn < h

8n: 8n

(2:6:27) (2:6:28)

khn kWp1 ( ) < C

holds, where C is the number from (2.6.1). Since the imbedding of Wp1 ( ) into C ( ) is continuous when p > 2 (see Theorem 1.6.2), the relations (2.6.26){(2.6.28) imply that, for any n, a number "n > 0 exists such that d("n hn ) Uad (2:6:29) where d("n hn ) = z j z 2 Wp1 ( ) kz ; hn kWp1 ( ) "n : (2:6:30) Hence, the functions hn de ned in (2.6.25) belong to U ad . From (2.6.22), (2.6.25), we easily deduce that hn ! h in Wp1 ( ). So, if the estimate (2.6.28) holds, then the lemma will be proved. 2. Let us establish the inequality (2.6.28). By (2.6.22), (2.6.25), we get
ZZ

@h p dx dy = Z Z @q p dx dy @x @x ZZ ZZ p @hn p dx dy: p @q dx dy = ln @x @x
ZZ

(2.6.31)

Similarly, we obtain

@h p dx dy @y

ZZ

@hn p dx dy: @y

(2:6:32)

2.6. Optimization of the spectrum

125

Introduce the function 0 1] 3 s ! f (s) =

ZZ

(jt + qjp ; jt + sqjp ) dx dy:

(2:6:33)

(Notice that the function (x y) ! s (x y) = t + sq(x y) is positive on for every s 2 0 1], so we could write below (t + sq)p instead of jt + sqjp .) From (2.6.22), (2.6.25), and (2.6.33), it follows that

f (ln ) =

ZZ

jhjp dx dy ;

ZZ

jhn jp dx dy:

(2:6:34)

The function s ! f (s) is di erentiable on the interval (0,1). Applying to it the mean value theorem, we get

f (s) = f (1) + @f ( )(s ; 1) @s

= ;(s ; 1)p
ZZ

jt + qjp;1 q dx dy
(x y) j (x y) 2 (x y) j (x y) 2

s 2 (0 1) 2 (s 1):

(2.6.35)

Let us denote
1= 2=

q (x y ) > 0 q (x y ) < 0 :
ZZ
2

(2.6.36) (2.6.37) (2:6:38) (2:6:39)

By (2.6.24), we have

a1 =
and Obviously, mes

ZZ
1

q dx dy > 0
ZZ

a2 =

q dx dy 0

q dx dy = a1 + a2 > 0:
on on

1 > 0 and

jt + qjp;1 > tp;1 0 < jt + qjp;1 < tp;1

Due to (2.6.36), (2.6.38), and (2.6.40), we obtain
ZZ
1

1 2

(2.6.40) (2.6.41) (2:6:42)

jt + qjp;1 q dx dy >

ZZ
1

tp;1 q dx dy = tp;1 a1 :

126

Chapter 2. Optimal control by coe cients in elliptic systems

ZZ
2

By virtue of (2.6.37), (2.6.38), and (2.6.41), we get

jt + qjp;1 q dx dy tp;1 a2 :

(2:6:43)

Further, by (2.6.36){(2.6.39), (2.6.42), and (2.6.43), we have

ZZ

jt + qjp;1 q dx dy =
+
ZZ
2

ZZ
1

jt + qjp;1 q dx dy 2 (0 1):

jt + qjp;1 q dx dy > tp;1 (a1 + a2 ) > 0

From here and from the relation (2.6.35) f (s) > 0 s 2 (0 1): (2:6:44) Now, using (2.6.31), (2.6.32), (2.6.34), and (2.6.44) and noticing that 0 < ln < 1, we obtain This inequality yields (2.6.28). The lemma is proved.

khkWp1( ) > khnkWp1 ( ) :

Finite-dimensional problem Let fHn g1 be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying n=1
the limit density condition, i.e., lim inf kh ; z kWp1 ( ) = 0 n!1 z2H
n

h 2 Wp1 ( ):

(2:6:45) (2.6.46) (2.6.47)

Consider the following problem: Find a function hn such that

One may choose a subsequence fhmgm=1 of the sequence fhng1 k such that hm ! n= h0 weakly in Wp1 ( ).
1

be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying the condition (2.6.45). Then, for any n su ciently large (say, n k), the problem (2.6.46), (2.6.47) has a solution. This problem has a solution for each n if the set Hn \ Uad is not empty for all n. Moreover, lim f (h ) = f (h0 ) = h2U f (h): inf (2:6:48) n!1 n
ad

Theorem 2.6.2 Let the conditions of Theorem 2.6.1 be ful lled and let a function h0 solving the problem (2.6.7) be not identically equal to h in . Also, let fHn g1 n=1

hn 2 Hn \ Uad f (hn ) = h2Hinf U f (h): \

n
ad

2.6. Optimization of the spectrum

127

Proof. 1. By virtue of Lemma 2.6.1, for any " > 0, there exists an element h" 2 U ad such that kh0 ; h" kWp1 ( ) ": (2:6:49) Since h" 2 U ad , there exists 2 R such that
0<

"

d( h" ) = h j h 2 Wp1 ( ) kh ; h" kWp1 ( ) : (2:6:51) From (2.6.45), it follows that, for any n su ciently large (say, n k), we can nd an element gn from Hn such that gn 2 d( h" ) moreover, due to (2.6.50), gn 2 Uad . By (2.6.49){(2.6.51), we have kgn ; h0 kWp1 ( ) kgn ; h" kWp1 ( ) + kh" ; h0 kWp1 ( ) 2": The above argument implies the existence of a sequence fgn g1 k such that n= gn 2 Hn \ Uad (2.6.52) 1 ( ): gn ! h0 in Wp (2.6.53) 2. So, the set Hn \ Uad is not empty when n k. It is easy to see that Hn \ Uad is a bounded and closed subset of Hn (Hn being endowed with the topology of Wp1 ( )). Since the set Hn is nite-dimensional, we conclude that Hn \ Uad is a

where

d( h" ) Uad

(2:6:50)

Due to (2.6.7), (2.6.47), (2.6.52), we get f (gn ) f (hn ) f (h0 ) 8n: From here, taking into account (2.6.55), we obtain (2.6.48). By virtue of (2.6.46), (2.6.48), the elements hn satisfy the conditions (2.6.9), (2.6.10). The argument from the proof of Theorem 2.6.1 implies the existence of a subsequence fhm g of the sequence fhn g such that hm ! h0 weakly in Wp1 ( ). Note that the problems (2.6.7) and (2.6.46), (2.6.47) may, in general, possess not a unique solution, so we may choose subsequences of the sequence fhng convergent to the solutions to the problem (2.6.7).

compact topological space. Theorem 2.5.2 and the relations (2.6.5), (2.6.6) yield that 9 h ! f (h) is a continuous mapping of Yp (endowed with the= (2:6:54) topology generated by the Wp1 ( )-weak one) into R when p > 2: Now, Theorem 1.3.9 implies the problem (2.6.46), (2.6.47) to have a solution for n k, and for any n provided the set Hn \ Uad is not empty for every n. By (2.6.53), (2.6.54), we have lim f (g ) = f (h0 ): (2:6:55) n!1 n

128

Chapter 2. Optimal control by coe cients in elliptic systems

(2.6.57), we may rewrite the problem (2.5.2) as (ui i ) 2 V R ui 6= 0 Bh ui = i ui : (2:6:61) Let Pm be the orthogonal projection of V onto Vm in the scalar product generated by ah . Then (see Lemma 1.5.1 and Remark 1.5.1), we may present the problem (2.6.56) in the form (uim im ) 2 Vm R uim 6= 0 Pm ( im uim ; Bh uim ) = 0:

Proof. Considering the form ah(u v) as a scalar product in V and noticing

Solution of the optimal control problem (2.6.7) and its nite-dimensional analogs| the problems (2.6.46), (2.6.47)|is connected with the computation of the eigenvalues of the problem (2.5.2). Let us show that these can be found by means of a passage to the limit, on the basis of the Galerkin approximations. Let fVm g1=1 be a sequence of nite-dimensional subspaces of V satisfying m the limit density condition (2.4.39). We consider the following eigenvalue problem in the space Vm : (uim im ) 2 Vm R uim 6= 0 v 2 Vm : (2.6.56) im ah (uim v ) = bh (uim v ) Here, uim , im are approximate eigenfunctions and eigenvalues of the problem (2.5.2). Since, for any h 2 Yp , the bilinear form ah (u v) is symmetric, continuous, and coercive on V V , we may consider it as a scalar product on V (see Remark 1.5.2). Because bh(u v) is a bilinear, symmetric, continuous form on U U and (2.5.1) is valid, we deduce from the Riesz theorem that bh(u v) = ah (Bh u v) Bh u 2 V u v 2 V: (2:6:57) In this case, Bh is a linear, compact, selfadjoint operator acting in V . g (2:6:58) Theorem 2.6.3 Let the conditions of Theorem 2.5.1 be satis ed and let fVmg be a sequence of nite-dimensional subspaces of V satisfying the condition (2.4.39). Also, let uim , im , i = 1 2 : : : km , be determined by a solution to the problem (2.6.56) (km being the dimension of the space Vm ). Then, the following estimate for the eigenvalue error holds j im ; i j 21=2 k(I ; Pm ) Bh kL(V V ) i = 1 2 : : : km (2:6:59) where i are the eigenvalues of the problem (2.5.2), I is the identity operator in V , Pm the orthogonal projection of V onto Vm . Moreover, lim k(I ; Pm ) Bh kL(V V ) = 0: (2:6:60) m!1

2.6.3 Computation of eigenvalues

2.6. Optimization of the spectrum

129

Taking into account that Pm u = u for any u 2 Vm , the latter problem can be rewritten as (uim im ) 2 V

uim 6= 0

im uim = (Pm

Bh Pm )uim : (2:6:62)

The operator Pm being selfadjoint, we deduce from (2.6.58) that Pm Bh Pm is a linear, compact, selfadjoint operator acting in V . Now, applying Theorem 1.5.6 to the operators Bh and Pm Bh Pm (to the problems (2.6.61) and (2.6.62)), we get

im ; i j

kBh ; Pm Bh Pm kL(V V )

i = 1 2 : : : km :

(2:6:63) (2:6:64)

For an arbitrary u 2 V , we set

w = (Bh ; Pm Bh Pm ) u:
We have

Pm w = Pm Bh P (m) u P (m) w = P (m) Bh u where P (m) = I ; Pm . Therefore, taking to notice that Bh P (m) is the adjoint operator of P (m) Bh , we get the estimates

kPm wkV kP (m) Bh kL(V V ) kukV kP (m)wkV kP (m) Bh kL(V V ) kukV (2.6.65) 2 = kP wk2 + kP (m) wk2 2kP (m) B k2 2: kwkV m V h L(V V ) kukV V
By (2.6.64) and (2.6.65), we have

kBh ; Pm Bh Pm kL(V V ) 21=2 kP (m) Bh kL(V V ) :

This inequality together with (2.6.63) implies the estimate (2.6.59). Denote the closed unit ball in V by Q. The operator Bh being compact, the set Bh (Q) is compact. Due to (2.4.39), the sequence of the operators fPm g converges simply to the operator I , that is,
m!1

lim kPm u ; ukV = 0

u 2 V:

Hence, taking into account that the set Bh (Q) is compact, by the Banach{Steinhaus theorem (see, e.g., Schwartz (1967)), we obtain that Pm ! I uniformly on Bh (Q), i.e., (2.6.60) holds. The theorem is proved. Note that the eigenvalue error estimates mentioned above, as well as more precise ones, may be found in Krasnoselsky et al. (1969), Michlin (1970), Gould (1966).

130

Chapter 2. Optimal control by coe cients in elliptic systems

2.7.1 Optimal control problem

2.7 Control under restrictions on the spectrum

Suppose we are given mappings Ai such that ! Ai is a continuous mapping of `1 0 into R, i = (2:7:1) 1 2 : : : k. De ne a set of admissible controls as Uad = h j h 2 Wp1 ( ) khkWp1 ( ) C h h ^ h (h) 0 Ai (h) 0 (i = 1 2 : : : k) : (2.7.2) Here, ^ ^ p > 2, C , h, h are positive numbers such that e1 < h < h < (2.7.3) e2 , where e1 , e2 are the positive numbers from (2.1.2) h ! (h) is a continuous mapping of Yp (endowed with the (2.7.4) topology generated by the Wp1 ( )-weak one) into R (h) = f 1(h) 2 (h) : : : g (2.7.5) i (h) = i being the eigenvalues of the problem (2.5.2). We suppose the goal functional to satisfy the following condition: h ! f (h) is a continuous mapping of Yp (equipped with (2:7:6) the topology generated by the Wp1 ( )-weak one) into R. The optimal control problem consists in nding a function h0 such that h0 2 Uad f (h0 ) = h2U f (h): inf (2:7:7)

Theorem 2.7.1 Let ah be a bilinear form on V V determined by the relations (2.1.1), (2.1.4){(2.1.7), and (2.1.10). Let also a mapping h ! bh satisfying the condition (2.5.12) be given. Assume that the nonempty set Uad is determined by the relations (2.7.1){(2.7.5), and the goal functional satis es the condition (2.7.6). Then, the problem (2.7.7) has a solution. Proof. Let fhng1=1 be a sequence such that n hn 2 Uad 8n (2.7.8) lim f (h ) = inf f (h): (2.7.9) n!1 n h2U
ad

ad

Due to (2.7.2), (2.7.3), the sequence fhng1 is bounded in Wp1 ( ) with n=1 p > 2. Hence, we may choose a subsequence fhm g1=1 such that m 1( ) hm ! z weakly in Wp hm ! z strongly in C ( ): (2:7:10)

2.7. Control under restrictions on the spectrum

131

From (2.7.8), (2.7.10), we deduce that

C lim inf khmkWp1 ( ) kz kWp1 ( m!1 ^ h z h C being the number from (2.7.2).

(2.7.11) (2.7.12)

Taking into account (2.7.4), we obtain from (2.7.8) and (2.7.10) that (z ) 0: (2:7:13)

By virtue of Theorem 2.5.2, (2.7.1), (2.7.8), and (2.7.10), we have

Ai (z ) 0

i = 1 2 : : : k:

Therefore, from (2.7.11){(2.7.13) we conclude that z 2 Uad. Finally, the relations (2.7.6), (2.7.9), (2.7.10) imply
m!1

lim f (hm ) = f (z ) = h2U f (h): inf

ad

Consequently, the function h0 = z is a solution to the problem (2.7.7).

2.7.2 Approximate solution of the problem (2.7.7)

hn 2 Hn \ Uad f (hn ) = h2Hinf U f (h): \
n
ad

Let fHn g1 be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying n=1 the condition (2.6.45). Consider the problem: Find a function hn such that (2:7:14)

We endow the set Uad de ned by the relations (2.7.1){(2.7.5) with the topology generated by the Wp1 ( )-one.
be a sequence of nite-dimensional subspaces of Wp1 ( ) meeting the condition (2.6.45). Also assume the existence of a sequence fqn g1 such that qn 2 U ad n=1 for each n and qn ! h0 in Wp1 ( ), h0 being a solution to the problem (2.7.7). Then, for every n su ciently large (say, for n l), the problem (2.7.14) has a solution. This problem has a solution for every n if the set Hn \ Uad is not empty for each n. Moreover,

Theorem 2.7.2 Let the conditions of Theorem 2.7.1 be satis ed and let fHng1=1 n

inf nlim f (hn ) = f (h0 ) = h2Uad f (h): !1

(2:7:15)

One may choose a subsequence fhmg of the sequence fhng such that hm ! h0 weakly in Wp1 ( ).

132

Chapter 2. Optimal control by coe cients in elliptic systems

U ad convergent to h0 in Wp1 ( ), we deduce, just as in the proof of Theorem 2.6.2, that there exists a sequence fgng1 l such that n= gn 2 Hn \ Uad (2.7.16) gn ! h0 in Wp1 ( ): (2.7.17) The set Hn \ Uad equipped with the topology induced by the Wp1 ( )-one is a compactum. Hence, (2.7.6) implies the problem (2.7.14) to be solvable for n l, and for each n if the set Hn \ Uad is not empty for all n.
By (2.7.6) and (2.7.17), we obtain lim f (g ) = f (h0 ): n!1 n (2:7:18)

Proof. Making use of the assumption about the existence of a sequence fqng

By (2.7.7), (2.7.14), (2.7.16), we have f (gn ) f (hn ) f (h0 ) whence, noticing (2.7.18), we obtain (2.7.15). Thus, the sequence fhng satis es the conditions (2.7.8), (2.7.9). Now, the argument of Theorem 2.7.1 yields the existence of a subsequence fhmg of the sequence fhng such that hm ! h0 weakly in Wp1 ( ).

In subsec. 2.7.2, we assumed the existence of a sequence of interior points of the set Uad convergent to the solution to the problem (2.7.7). However, the veri cation of this assumption is rather di cult. This is why we consider another approach that makes no use of that condition. Let t 2 R, 1 t eh2 (e2 being the number from (2.1.2)). De ne a set ^ ^ U (t) = h j h 2 Wp1 ( ) khkWp1 ( ) tC h h th (h) t ; 1 Ai (h) t ; 1 (i = 1 2 : : : k) : (2.7.19)

The sequence fqng

2.7.3 Second method of approximate solution of the problem (2.7.7)

Here,

Supposing the conditions of Theorem 2.7.1 to be ful lled, we consider the ^ problem: For a xed t such that t > 1, th e2 (e2 being the positive number from (2.1.2)), nd a function ht satisfying ht 2 U (t) f (ht ) = inf(t) f (h): (2:7:21)
h2U

C h ^ are the positive numbers from (2.7.2) p > 2: h (2:7:20) (t) , Uad = U (1) . Combining (2.7.2) and (2.7.19), we see that Uad U

2.7. Control under restrictions on the spectrum

133

Uad U (t) implies

Due to Theorem 2.7.1, the problem (2.7.21) has a solution, and the imbedding

where

h0 being the solution to the problem (2.7.7). De ne functions qn by qn = tn h0

1 < tn < t By (2.7.23) and (2.7.24),

f (ht ) f (h0 )

(2:7:22) (2:7:23) (2:7:24) (2:7:25)

lim t = 1: n!1 n

qn ! h0 in Wp1 ( ):

We endow the set U (t) with the topology induced by the Wp1 ( )-one. Let us show that qn 2 U (t) n l (2:7:26) where l is a positive number. Since h0 2 Uad , from (2.7.2), (2.7.23), and (2.7.24), we get

kqn kWp1 ( ) = tn khokWp1 ( ) < tkho kWp1 (

^ h h0 < tn h0 = qn < th

8n

tC

8n

(2.7.27) (2.7.28)

^ C , h, and h being the numbers from (2.7.2). By using (2.7.2), (2.7.4), and (2.7.25, we infer the following estimate for n su ciently large (say, n l) (qn ) < t ; 1 n l: (2:7:29) By (2.7.1) and Theorem 2.5.2, we obtain that

h ! Ai (h) is a continuous mapping of Yp (endowed with= the topology generated by the Wp1 ( )-weak one, p > 2) into (2:7:30) R (i = 1 2 : : : k). Since Ai (h0 ) 0 and t > 1, the formulas (2.7.25) and (2.7.30) yield the following inequality to hold for l su ciently large: Ai (qn ) < t ; 1 n l i = 1 2 : : : k: (2:7:31) Due to (2.7.4), (2.7.29){(2.7.31), for any n l, there exists "n > 0 such that (h) t ; 1 h 2 d("n qn ) n l (2.7.32) Ai (h) t ; 1 h 2 d("n qn ) n l i = 1 2 : : : k (2.7.33) 1 ( ) kh ; q k 1 d("n qn ) = h j h 2 Wp (2.7.34) n Wp ( ) "n :

134

Chapter 2. Optimal control by coe cients in elliptic systems

The imbedding Wp1 ( ) ! C ( ) being continuous for p > 2, from (2.7.27), (2.7.28), we deduce the existence of a ball d( n qn ), n "n , de ned by the relation (2.7.34) such that the following estimate holds: ^ khkWp1( ) tC h h th h 2 d( n qn ) n l: (2:7:35) Now, (2.7.19), (2.7.27), (2.7.32), (2.7.33), and (2.7.35) imply (2.7.26).

Approximate solution of the problem (2.7.7) Let fHn g1 be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying n=1 the condition (2.6.45) and Hn Hn+1 8n: (2:7:36)
Consider the problem: Find a function hn such that

hn 2 Hn \ U (t) f (hn ) = inf (t) f (h):

h2Hn \U

(2.7.37) (2.7.38)

A subsequence fhmg of the sequence fhn g may be chosen such that hm ! g weakly in Wp1 ( ), g 2 U (t) , and
m!1

^ U (t) be determined by the relations (2.7.19), (2.7.20) when t > 1, th e2 (e2 being the positive number from (2.1.2)). Further, let fHng be a sequence of nitedimensional subspaces of Wp1 ( ) meeting the conditions (2.6.45), (2.7.36) and let h0 , ht be solutions to the problems (2.7.7), (2.7.21), respectively. Assume that the set H1 \ U (t) is not empty. Then, for each n, the problem (2.7.37), (2.7.38) has a solution such that f (ht ) nlim f (hn ) f (h0 ): (2:7:39) !1 lim f (hm ) = nlim (hn ) = f (g): !1 (2:7:40)

Theorem 2.7.3 Let the conditions of Theorem 2.7.1 be ful lled and let a set

Proof. By the hypothesis, the set H1 \ U (t) is not empty. Hence, (2.7.36) implies the set Hn \ U (t) to be nonempty for each n. Taking into account (2.7.4) and (2.7.30), we easily see that the set Hn \ U (t) is a compact topological space with the topology generated by the Wp1 ( )-one. Now, (2.7.6) implies the solvability of the problem (2.7.37), (2.7.38) for each n. Combining (2.7.21), (2.7.38), we get
f (ht ) f (hn )
From (2.7.36) and (2.7.38)

8n: 8n:

(2:7:41)

f (hn+1 ) f (hn )

2.7. Control under restrictions on the spectrum

135

Therefore, by (2.7.41), we conclude the number sequence ff (hn )g1 to converge n=1 and lim f (h ) f (ht ): (2:7:42) n!1 n Using (2.6.45), (2.7.25), and (2.7.26), analogously to the proof of Theorem 2.6.2, one shows the existence of a sequence fwn g1 such that n=1

wn 2 Hn \ U (t) wn ! h0 in Wp1 ( ):

(2.7.43) (2.7.44) (2:7:45) (2:7:46)

By (2.7.6) and (2.7.44), we have

n!1

lim f (wn ) = f (h0 ):

Due to (2.7.38) and (2.7.43), we get

f (wn ) f (hn ):
that

fhn g1=1 to be bounded in Wp1 ( ). Let us choose a subsequence fhmg1=1 such n m

hm ! g hm ! g
weakly in Wp1 ( ) strongly in C ( ): (2.7.47) (2.7.48) (2.7.49) (2.7.50) (2:7:51) (2:7:52)

Now, (2.7.42), (2.7.45), and (2.7.46) yield (2.7.39). (2.7.37) implies the sequence

By (2.7.4), (2.7.19), (2.7.30), (2.7.37), and (2.7.47), we get (g) = mlim (hm ) t ; 1 !1 Ai (g) = mlim Ai (hm ) t ; 1 !1 The relations (2.7.19), (2.7.37), and (2.7.48) imply ^ h g th: By virtue of (2.7.19), (2.7.37), and (2.7.47), we obtain

i = 1 2 : : : k:

tC lim inf khmkWp1 ( m!1

kgkWp1 ( ) :

(2.7.49){(2.7.52) yield g 2 U (t) . Finally, by (2.7.6), (2.7.47), we get (2.7.40), concluding the proof.

Remark 2.7.1 The function g and the functions hn solving the problem (2.7.37),
(2.7.38) belong to U (t) and, generally speaking, they do not belong to Uad. But since the parameter t can be chosen as near to 1 as desired, \the getting of the functions hn and g out of the region Uad" can be made arbitrarily small.

136

Chapter 2. Optimal control by coe cients in elliptic systems

Let A be a continuous functional de ned on the space `1 0 . The functional A is said to satisfy at a point 2 `1 0 the condition with integer indices j1 j2 : : : jm if there exists a neighborhood d( ) of the point in `1 0 such that A = f ( j1 j2 : : : jm ) = f j g1 2 d( ) j =1 f being a continuously di erentiable function de ned on an open set Q in Rm such that ( j1 j2 : : : jm ) 2 Q 2 d( ): We endow the set G de ned in (2.2.44) with the topology generated by the Wp1 ( )-one. Provided the hypothesis of Theorem 2.5.2 is ful lled, de ne a functional on G by (h) = A (h): (2:7:53) From the continuity of the functional A and Theorem 2.5.2, we conclude to be a continuous functional on G. The following theorem on the di erentiability of this functional holds . Theorem 2.7.4 Let the conditions of Theorem 2.5.1 be satis ed and let (2.2.52), (2.5.12) hold. Also, let h ! bh be a Frechet continuously di erentiable mapping of G into N2 . ~ ~ Assume that h 2 G and a continuous functional A meets at the point (h) the ~ ) j2 (h) : : : jm (h) are ~ ~ condition with integer indices j1 j2 : : : jm , and j1 (h ~ simple nonzero eigenvalues. Then, there exists a neighborhood d(h) of the point ~ h in G such that the functional de ned by (2.7.53) is Frechet continuously ~ ~ di erentiable in d(h) and its Frechet derivative 0 (h) at a point h 2 d(h) is given by the formula
0

Theorem on the di erentiation of functionals

2.7.4 Di erentiation of the functionals h ! Ai (h) and necessary conditions of optimality

(h)q =

q 2 Wp1 ( ) (2.7.54) b0h , a0h being the Frechet derivatives of the functions h ! ah and h ! bh at point h, ujs (h) the eigenfunctions of the problem (2.5.2) corresponding to the eigenvalues js (h) and satisfying the condition ah (ujs (h) ujs (h)) = 1. Proof. The inequality (2.2.57) and the argument below it yield h ! ah to be a Frechet continuously di erentiable mapping of G into L2 (V V R), and its Frechet derivative a0h at point h 2 G to be given by the relation ZZ X k da ij a0h q(u v) = dt (h)q(Pi u)(Pj v) dx dy i j =1

m X @f (h) ; ~ j1 (h) j2 (h) : : : jm (h) s=1 @ js (b0hq ; js (h)a0h q) (ujs (h) ujs (h))

2.7. Control under restrictions on the spectrum

137

q 2 Wp1 ( ) u v 2 V:
~ Now, Theorem 1.11.3 implies the existence of a neighborhood d1 (h) of the ~ in G such that the functionals h ! js (h), s = 1 2 : : : m, are Frechet point h ~ continuously di erentiable in d1 (h) and their Frechet derivatives 0js (h) at a point ~ ) are given by h 2 d1 (h
0 0 js (h)q = (bh q ; js (h)ah q ) (ujs (h) ujs (h)) 0

q 2 Wp1 ( ) s = 1 2 : : : m:

Further, the well-known result on the di erentiation of a composite function ~ (see subsec. 1.9.1, Property 4) yields the existence of a neighborhood d(h) of the ~ point h in G such that the functional from (2.7.53) is Frechet continuously ~ ~ di erentiable in d(h), and its Frechet derivatives 0 (h) at a point h 2 d(h) is given by the formula (2.7.54). By using Theorem 2.7.4, one can nd the Frechet derivatives of the functionals h ! Ai (h) when this theorem is applicable. These derivatives may be used when one solves the nite-dimensional problem (2.7.14). ^ The restrictions h h h lead to the functionals '1 and '2 de ned by (2.2.33) which are not Frechet di erentiable. As noted in subsec. 2.2.3, by introducing a ^ corresponding "-net, one can approximate the restrictions h h h by those of the form Qi (h) 0, as precisely as desired (Qi being the Frechet di erentiable functionals determined by the formulas (2.2.37)). Now, replace the set Uad from (2.7.2) by the following one ~ Uad = h j h 2 Wp1 ( ) khkWp1 ( ) C Qi (h) 0 i = 1 2 : : : 2r (h) 0 Aj (h) 0 j = 1 2 : : : k : (2.7.55) that Given a goal functional f , consider the problem of nding a function h0 such ~ h0 2 Uad

Restatement of the problem and necessary conditions of optimality

f (h0 ) = inf f (h): ~

h2Uad

(2:7:56)

Using the argument from the proof of Theorem 2.7.1, we establish the following fact.

Theorem 2.7.5 Let ah be a bilinear form on V V de ned by the relations (2.1.1), (2.1.4), (2.1.6), (2.1.7), (2.1.10), (2.2.52). Assume that h ! bh is a continuously Frechet di erentiable mapping of G into N2 , and (2.5.12) holds. Suppose that a ~ nonempty set Uad is determined by the relations (2.7.55), (2.2.37), (2.7.1), (2.7.3), (2.7.4), and the goal functional f meets the condition (2.7.6). Then, there exists a solution h0 to the problem (2.7.56).

138

Chapter 2. Optimal control by coe cients in elliptic systems

Let J = (1 2 : : : k) and let J0 = (i1 i2 : : : il ) be a subset of J such that

Aj (h0 ) = 0 Aj (h0 ) < 0

Introduce the following notations:

j 2 J0 j 2 J n J0 :

g0 (h) = f (h) g1 (h) = khkWp1( ) ; C gi (h) = Qi;1 (h) i = 2 3 : : : 2r + 1 g2r+2(h) = (h) g2r+2+s (h) = Ais (h) s = 1 2 : : : l is 2 J0
where h 2 G.

(2.7.57)

Theorem 2.7.6 Let the conditions of Theorem 2.7.5 be ful lled and let d(h0 ) be some neighborhood of the function h0 in G. Suppose that the functionals f and are Frechet di erentiable on d(h0 ). Also assume the functionals Aj , j 2 J0 , meet at a point (h0 ) the condition with indices i1(j ) i2(j ) : : : im(j) (j ), and the corresponding eigenvalues i1 (j) (h0 ) i2 (j) (h0 ) : : : im(j) (j) (h0 ) are simple ones and do not equal to zero. Then, there exist real numbers 0 1 : : : 2r+l+2 that do not vanish simultaneously and such that
0

0 :::

2r+l+2

2rX +l+2

i gi (h0 ) = 0

i = 1 2 : : : 2r + l + 2

i=0

0 i gi (h0 ) = 0

gi0 (h0 ) being the Frechet derivatives of the functionals gi at a point h0 (see (2.7.57)). If there exists an element v 2 Wp1 ( ) such that gi0 (h0 )v < 0 for i = 1 2 : : : 2r + l + 2, then 0 6= 0 and we can take 0 = 1.

Proof. By virtue of Theorem 2.7.4, the functionals g2r+2+s, s = 1 2 : : : l, de ned in (2.7.57) are continuously Frechet di erentiable in some neighborhood of the point h0 in G. The functional g1 is also continuously di erentiable in this neighborhood (see Theorem 1.10.1). From (2.2.37), (2.7.57), and the imbedding theorem we have g2 g3 : : : g2r+1 to be a ne continuous functionals in Wp1 ( ), so they are continuously di erentiable in G. Since by the hypothesis of the theorem the functionals f and are Frechet di erentiable in d(h0 ), due to (2.7.57) the functionals g0 and g2r+2 are Frechet di erentiable in d(h0 ). Next, use Theorem 1.12.1. Set X = Wp1 ( ), U = G, Y = f0g R, and let Fh = 0 for all h 2 G. Then, Theorem 2.7.6 is a consequence of Theorem 1.12.1.

2.8. The basic optimal control problem

139

2.8 The basic optimal control problem

Let the set of admissible controls be of the form

2.8.1 Setting of the problem. Existence theorem

Uad = h h 2 Wp1 ( ) khkWp1 (
Here
)

^ C h h h

k (h)

0 (k = 1 2 : : : q) : (2.8.1) (2.8.2) (2.8.3) (2:8:4) (2:8:5) (2:8:6)

^ ^ p > 2, C , h, h are positive numbers such that e1 < h < h < e2 , where e1 , e2 are the positive numbers from (2.1.2) 9 h ! k (h) is a continuous mapping of Yp (equipped with= the topology induced by the Wp1 ( )-weak one) into R when p > 2 (k = 1 2 : : : q). We suppose that the set Uad is not empty. Let f be a goal functional satisfying the following condition: h ! f (h) is a continuous mapping of Yp (equipped with the topology induced by the Wp1 ( )-weak one) into R.

The optimal control problem consists in nding a function h0 such that

h0 2 Uad

f (h0 ) = h2U f (h): inf

ad

Theorem 2.8.1 Let a set Uad be determined by the relations (2.8.1){(2.8.4), and Proof. In virtue of (2.8.4), there exists a sequence fhng such that hn 2 Uad 8n
n!1

let a goal functional satisfy the condition (2.8.5). Then, there exists a solution to the problem (2.8.6).

lim f (hn ) = h2U f (h): inf ad

(2.8.7) (2.8.8)

By the de nition of the set Uad , the sequence fhng is bounded in Wp1 ( ). So, we can choose a subsequence fhmg such that

hm ! g

weakly in Wp1 ( )
)

(2:8:9)
)

g being an element of Wp1 ( ). Then C lim inf khmkWp1 ( m!1

kgkWp1 (

(2:8:10)

140

Chapter 2. Optimal control by coe cients in elliptic systems

C being the constant from (2.8.1). p > 2, we conclude that

By using (2.8.9) and the compactness of the embedding Wp1 ( ) ! C ( ) as ^ h g h: (2:8:11) (2:8:12)

So, the function h0 = g is a solution to the problem (2.8.6).

The formulas (2.8.10){(2.8.12) imply that g 2 Uad , and from (2.8.5), (2.8.8), and (2.8.9), we deduce f (g) = mlim f (hm ) = h2U f (h): inf !1
ad

Further, the relations (2.8.1), (2.8.3), (2.8.7), and (2.8.9) yield lim (h ) = k (g) 0 k = 1 2 : : : q: m!1 k m

2.8.2 Approximate solution of the problem (2.8.6)

Expansion of the set Uad Let t 2 R and 1 t min(2 ; eh1
(2.1.2). De ne a set
e2 ), where e1 and e2 are the numbers from ^ h

U (t) = h j h 2 Wp1 ( ) khkWp1 (

^ tC h(2 ; t) h th k (h) t ; 1 (k = 1 2 : : : q ) : (2.8.13)

Here, k (h) are the functionals de ned in subsec. 2.8.1. Comparing (2.8.1) and (2.8.13) gives Uad U (t) Uad = U (1) and the argument from the proof of Theorem 2.8.1 implies the following statement. Theorem 2.8.2 Let a set U (t) be determined by the formulas (2.8.13), (2.8.2), (2.8.3), and (2:8:14) 1 < t min 2 ; e1 e2 : ^
and let the set be nonempty. Let a goal functional f satisfy the condition (2.8.5). Then, there exists a function ht such that ht 2 U (t) f (ht ) = inf(t) f (h): (2:8:15)

U (t)

h h

Remark 2.8.1 We introduce the right hand side of the estimate (2.8.14) to make the inequality e1 h e2 hold for all h 2 U (t) , e1 and e2 being the positive
numbers from (2.1.2), i.e., in order to get the inclusion U (t) Yp , which is needed because the functionals k and f are de ned on Yp .

h 2U

2.8. The basic optimal control problem

141

Endow the set U (t) with the topology induced by the strong topology of the space Wp1 ( ). The following statement is valid.

Lemma 2.8.1 Let a set Uad be determined through the relations (2.8.1){(2.8.4),
and a set U (t) through the relations (2.8.13), (2.8.14), (2.8.2), and (2.8.3). Then, Uad U (t) .

Proof. Let us show that g 2 U (t) for an arbitrary element g from Uad. For any h 2 Wp1 ( ), we have khkWp1 ( ) kgkWp1( ) + kh ; gkWp1 ( ) : (2:8:16) Since g 2 Uad , we have kgkWp1( ) C , where C is the constant from (2.8.1). Hence,
(2.8.16) implies the existence of a number
1 > 0 such that

khkWp1 (
where

tC

h 2 d( 1 g)
) 1

(2:8:17)

As the imbedding Wp1 ( ) ! C ( ) is continuous for p > 2 and h exists a number 2 > 0 such that ^ h(2 ; t) h th h 2 d( 2 g):

d( 1 g) = h j h 2 Wp1 ( ) kh ; gkWp1 (

: g
^ h, there (2:8:18)

The function h ! k (h) is continuous in the topology generated by the Wp1 ( )-weak one, so that it is continuous in the topology generated by the strong one of Wp1 ( ). Thus, from the estimate k (g) 0, we deduce the existence of a number 3 > 0 such that
k (h)

t;1

h 2 d( 3 g):

(2:8:19)

Setting = min( 1 2 3 ), taking into account (2.8.17){(2.8.19), we get d( g) U (t). Hence, g 2 U (t) and the lemma is proved.

conditions (2.6.45) and (2.7.36). Let us consider the problem of nding a function hn such that

Approximate solution of the basic problem Let fHn g be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying the
hn 2 Hn \ U (t) f (hn ) = inf (t) f (h):
h2Hn \U

(2.8.20) (2.8.21)

142

Chapter 2. Optimal control by coe cients in elliptic systems

Theorem 2.8.3 Let the assumptions of Theorems 2.8.1 and 2.8.2 be ful lled, let fHn g be a sequence of nite-dimensional subspaces of Wp1 ( ) satisfying the conditions (2.6.45) and (2.7.36), and let h0 , ht be solutions to the problems (2.8.6), (2.8.15), respectively. Suppose that the set H1 \ U (t) is not empty. Then, for any
~ One can choose from the sequence fhn g a subsequence fhmg such that hm ! h 1 ( ), h 2 U (t) , and ~ weakly in Wp ~ lim f (h ) = lim f (h ) = f (h): (2:8:23) m!1 m n!1 n compact topological space endowed with the topology generated by the topology of

n, the problem (2.8.20), (2.8.21) has a solution hn such that f (ht ) nlim f (hn ) f (h0 ): !1

(2:8:22)

Proof. The set H1 \ U (t) is not empty by the hypothesis. From here and from (2.7.36), we deduce Hn \ U (t) to be nonempty for all n. The set Hn \ U (t) is a

Wp1 ( ). This conclusion together with (2.8.5) yields the problem (2.8.20), (2.8.21) to be solvable for any n. The relations (2.8.15) and (2.8.21) imply f (ht ) f (hn ) 8n: (2:8:24)
From (2.7.36) and (2.8.21), we get f (hn+1 ) f (hn ): Combining this inequality with (2.8.24), we deduce the sequence of numbers ff (hn)g1=1 to be convergent and n lim f (h ) f (ht ): (2:8:25) n!1 n Due to Lemma 2.8.1, h0 2 U (t) . From here and from (2.6.45), we derive the existence of a sequence fgng such that gn 2 Hn \ U (t) lim kg ; h0 kWp1 ( ) = 0: (2:8:26) n!1 n
n!1

lim f (gn ) = f (h0 ): (2:8:27) The relations (2.8.20), (2.8.21), and (2.8.26) yield f (gn ) f (hn ): Therefore, by (2.8.25) and (2.8.27), we get (2.8.22). By (2.8.13) and (2.8.20), we see that the sequence fhn g is bounded in Wp1 ( ). Choose a subsequence fhmg from it such that ~ hm ! h weakly in Wp1 ( ) (2.8.28)

Hence,

2.9. The combined problem

143 (2.8.29)

~ hm ! h

strongly in C ( ):

Repeating the above argument and making use of (2.8.3), (2.8.13), (2.8.20), ~ (2.8.28), and (2.8.29), we easily see that h 2 U (t) . Now, (2.8.5) and (2.8.28) yield (2.8.23), concluding the proof.

2.9 The combined problem

Let us consider a combined problem as an application of the basic problem. Suppose that
9 u v ! a(i) (u v) is a bilinear form on V V depending= h on parameter h 2 Yp that is determined by the relations (2.1.1), (2.1.4){(2.1.7), and (2.1.10) (i = 1 2 3) u v ! bh(u v) is a bilinear, symmetric, continuous form on X1 X1 depending on parameter h 2 Yp u v ! Gh w (u v) is a bilinear, symmetric, continuous form on X2 X2 depending on parameter (h w) 2 Yp V .

(2.9.1) (2.9.2) (2.9.3)

Here

Xi is a Banach space (i = 1 2), V V ! Xi is compact.

Xi , the imbedding

(2:9:4)

We assume also that 9 h ! bh is a continuous mapping of Yp (endowed with the> = topology induced by the Wp1 ( )-weak one) into> L2 (X1 X1 R) 9 h w ! Gh w is a continuous mapping of Yp V (equipped> = with the topology generated by the product of the Wp1 ( )-> weak topology and of the V -strong one) into L2 (X2 X2 R).

(2.9.5) (2.9.6)

Fix an element f1 from V . For a given element h 2 Yp , consider the following three problems. First problem: Find a function uh such that

uh 2 V

a(1) (uh v) = (f1 v) h

v 2 V: v 2 V:

Second problem (the eigenvalue problem): Find ( i ui ) 2 R V such that

(2) i ah (ui v ) = bh (ui v )

144

Chapter 2. Optimal control by coe cients in elliptic systems

Third problem: Find ( i ui ) 2 R V such that (3) v2V i ah (ui v ) = Gh uh (ui v ) uh being a solution to the rst problem. We are going to set control by the function h. So, let us de ne a set of admissible controls by ^ Uad = h h 2 Wp1 ( ) khkWp1 ( ) C h h h (2.9.7) 1 (h uh ) 0 2 ( ) 0 3 ( ) 0 : Here, ^ p > 2, C , h, h are positive numbers such that e1 < h < ^ < h (2.9.8) e2 , where e1 and e2 are the positive numbers from (2.1.2) 9 h u ! 1 (h u) is a continuous mapping of Yp V (equipped= with the topology generated by the product of the Wp1 ( )(2.9.9) weak topology and of the V -strong one) into R (2.9.10) 2 , 3 are continuous mappings of the space `1 0 into R, g = f 1 2 3 : : : g and = f 1 2 3 : : : g are the eigenvalues of the second and third problems, respectively. The function uh appearing as an argument of functional 1 (see (2.9.7)) is a solution to the rst problem for a xed f1 2 V . About the goal functional f we suppose that h ! f (h) is a continuous mapping of Yp (endowed with the (2:9:11) topology induced by the Wp1 ( )-weak one) into R. Consider the following optimal control problem: Find a function h0 such that h0 2 Uad f (h0 ) = h2U f (h): inf (2:9:12)
ad

Notice that the problem (2.9.12) is an abstract analog of the problem of optimization of a plate (shell) under restrictions on the strength, stability, and free oscillation frequencies (cf. subsec. 5.3.4). Theorem 2.9.1 Let the relations (2.9.1){(2.9.6) be ful lled and let a nonempty set Uad be de ned by the relations (2.9.7){(2.9.10). Assume that the goal functional satis es the condition (2.9.11). Then, the problem (2.9.12) has a solution. Proof. De ne a functional 1 (h) as follows h ! 1 (h) = 1(h uh ) where uh is a solution to the rst problem. By (2.9.9) and Lemma 2.1.1, we conclude the mapping h ! 1 (h) to satisfy the condition (2.8.3). Further, introduce a functional 2 (h) h ! 2 (h) = 2 ( (h))

2.10. Optimal control problem for the case when the state of the system

:::

145

where (h) is the set of the eigenvalues of the second problem, depending on parameter h 2 Yp . Combining (2.9.2), (2.9.5), (2.9.10), and Theorem 2.5.2, we obtain that the mapping h ! 2 (h) meets the condition (2.8.3). De ne a functional h ! 3 (h) = 3 ( (h)) where (h) is the set of the eigenvalues of the third problem. Using Lemma 2.1.1, Theorem 2.5.2, and the relations (2.9.3), (2.9.6), (2.9.10), it is easy to see that the mapping h ! 3 (h) satis es the condition (2.8.3). With the notations introduced, the problem (2.9.12) is reduced to the problem (2.8.6). Now, Theorem 2.8.1 implies that (2.9.12) has a solution. It should be noted that, provided the set Uad is determined by the relations (2.9.7){(2.9.10), an approximate solution to the problem (2.9.12) may be constructed by using the technique of subsec. 2.8.2.

2.10 Optimal control problem for the case when the state of the system is characterized by a set of functions
Untill to now we studied optimal control problems when the control h belonged to the set Yp determined by the relations (2.1.2) and (2.1.3), and the state of the system uh was a solution to the problem (2.1.12). The function f in (2.1.12) was supposed to be xed and given. In this case, to every control h 2 Yp there corresponds a function uh 2 V that determines the state of the system. We will now study control problems when the function f 2 V is not xed, and instead of f we will be given some set T V . This problem is an abstract analog of the problem of optimization of a construction for a class of loads (T corresponding to this class). Provided a form ah is determined by the relations (2.1.1), (2.1.4){(2.1.6) and the inequalities (2.1.7) and (2.1.10) hold, for any element h 2 Yp and for any function g 2 V , there exists a unique function u(h g) such that

2.10.1 Setting of the problem

u(h g) 2 V

ah (u(h g) v) = (g v)

v2V

(2:10:1)

Now, for every control h 2 Yp , the state of the system Qh coincides with the set of functions u(h g), where g runs through the set T , i.e.,

Qh = fu(h g)gg2T :
Here, one can reason in another way. Since, for a given element h 2 Yp , the bilinear form ah is symmetric, continuous, and coercive on V V , the Riesz theorem implies

146

Chapter 2. Optimal control by coe cients in elliptic systems

that problem (2.10.1) has a unique solution u(h g) and well-de ned is an operator Nh 2 L(V V ) such that Nhg = u(h g): (2:10:2) Then, for a given element h 2 Yp , the state of the system is the image of the set T under the mapping Nh , i.e.,

Qh = Nh fT g = fu(h g)gg2T :
Introduce the notation

(2:10:3) (2:10:4)

B = h j h 2 C ( ) e1 h e2

e1 , e2 being the positive numbers from (2.1.2). By virtue of the imbedding theorem, Yp B when p > 2 (the set Yp de ned by (2.1.2)). Let mappings Ai : B V V ! R be de ned such that 9 h u g ! Ai (h u g) is a continuous mapping of B V V = (equipped with the topology generated by the product of (2:10:5) the topologies of C ( ), V , V ) into R (i = 1 2 : : : q). De ne mappings ;i : B V ! R as follows: h g ! ;i (h g) = Ai (h u(h g) g): (2:10:6)
Here, i = 1 2 : : : q, and u(h g) is a solution to the problem (2.10.1). Further, assume that T is a compact set in V (2:10:7) and the functions i : B ! R have the form

h ! i (h) = sup ;i (h g)
g2T
)

i = 1 2 : : : q:
i (h)

(2:10:8)

De ne a set of admissible controls Uad as

Uad = h j h 2 Wp1 ( ) khkWp1(

Here,

C h h ^ h

0 (i = 1 2 : : : q) : (2.10.9) (2:10:10)

^ p > 2, C , h, h are positive numbers such that e1 < h < ^ < h e2 , where e1 , e2 are the positive numbers from (2.1.2). Introduce a goal functional h ! f (h) such that h ! f (h) is a continuous mapping of Yp (endowed with the topology generated by the Wp1 ( )-weak one) into R.

(2:10:11)

2.10. Optimal control problem for the case when the state of the system

:::

147

The optimal control problem consists in nding a function h0 such that h0 2 Uad f (h0 ) = h2U f (h): inf (2:10:12)
ad

Notice that the basic di erence between the optimal control problem (2.10.12) and the problems investigated in Section 2.2 is that now the state of the system is determined not by the single function uh , but by the set Qh (see (2.10.3)), which is not, in general, a countable one, and that the set of admissible controls Uad is determined by Qh .

Theorem 2.10.1 Let ah be a bilinear form on V V determined by the relations (2.1.1), (2.1.4){(2.1.7), and (2.1.10). Suppose that the state of the system Qh is de ned by the formulas (2.10.1){(2.10.3), and the set of addmissible controls Uad is de ned by the relations (2.10.4){(2.10.10) and is nonempty. Let also the goal functional f (h) meet the condition (2.10.11). Then, the problem (2.10.12) has a solution. To prove Theorem 2.10.1 we will need two lemmas. Lemma 2.10.1 Let ah be a bilinear form on V V determined by the relations (2.1.1), (2.1.4){(2.1.7), and (2.1.10). Then, the function h g ! u(h g) de ned by the solution to the problem (2.10.1) is a continuous mapping of B V (endowed with the topology generated by the product of the topologies of the spaces C ( ) and V ) into V . Proof. Let fhn gng1=1 B V and n hn ! h0 in C ( ) (2.10.13) gn ! g0 in V : (2.10.14) Then, (h0 g0 ) 2 B V , and we need to prove that u(hn gn) ! u(h0 g0) in V as n ! 1: (2:10:15) Taking into account (2.1.1), (2.1.4), and (2.1.5), we obtain jahn (u u) ; ah0 (u u)j n ckuk2 u 2 V: (2:10:16) V Here, c is a positive number and (2:10:17) n = max max jaij (hn ) ; aij (h0 )j: ij
(x y)2

2.10.2 The existence theorem

Combining (2.1.5), (2.10.13), and Theorem 1.3.10, we deduce that lim = 0: n!1 n

(2:10:18)

148

Chapter 2. Optimal control by coe cients in elliptic systems

For all h 2 B , the bilinear form ah satis es the following condition (cf. subsec. 1.5.2) ah (u v) = (Ah u v) u v2V (2:10:19) where Ah 2 L(V V ), and the problem (2.10.1) is equivalent to the following one: u(h g) 2 V Ah u(h g) = g: (2:10:20) Due to (2.1.9), the operator Ah is selfadjoint, i.e., (Ah u v) = (Ah v u) u v2V hence, kAh kL(V V ) = sup j(Ah u u)j: Thus, by (2.10.16), (2.10.18), and (2.10.19), we conclude that nlim kAhn ; Ah0 kL(V V ) = 0: !1
kukV =1

(2:10:21)

We will denote by U the set of the invertible elements of the space L(V V ) endowed with the topology generated by that of the space L(V V ). The form ah being symmetric, continuous, and coercive on V V for all h 2 B , the Riesz theorem and (2.10.19) imply that

Ah 2 U

h 2 B:

By (2.10.13) and (2.10.21), we have h ! Ah is a continuous mapping of B into U : (2:10:22) Combining (2.10.13), (2.10.14), (2.10.20), and (2.10.22) and applying Theorem 1.8.1, we get (2.10.15). The lemma is proved. Introduce a set YpC = h j h 2 Wp1 ( ) khkWp1( ) C e1 h e2 (2:10:23)

e1 , e2 , and C being the positive numbers from (2.1.2) and (2.10.9). Lemma 2.10.2 Let ah be a bilinear form on V V determined by the relations (2.1.1), (2.1.4){(2.1.7), and (2.1.10). Then, the function h ! i (h), i = 1 2 : : : q, de ned by the formulas (2.10.5){(2.10.8), where u(h g) is a solution to the problem (2.10.1), is a continuous mapping of YpC (endowed with the topology generated by the Wp1 ( )-weak one ) into R when p > 2.

Proof. By Lemma 2.10.1 and the relations (2.10.5), (2.10.6), we obtain 9 h g ! ;i (h g) is a continuous mapping of B V (equipped= with the topology induced by the product of the topologies (2:10:24) of C ( ) and V ) into R.

2.10. Optimal control problem for the case when the state of the system

:::

149

Denote the closure of YpC in C ( ) as Y pC . The imbedding Wp1 ( ) ! C ( ) being compact when p > 2, by (2.10.4) and (2.10.23), we deduce that

Y pC is a compact set in B (with respect to the topology (2:10:25) generated by the C ( )-one). Now, the theorem on the continuity of the maximum function (Theorem 1.4.4) together with (2.10.7), (2.10.8), (2.10.24), and (2.10.25) yields that
into R. Let fhng be a sequence such that fhn g YpC hn ! h weakly in Wp1 ( ) p > 2: Then, h 2 YpC and hn ! h strongly in C ( ): Therefore, from (2.10.26), we deduce that i = 1 2 ::: q nlim i (hn ) = i (h) !1

h ! i (h) = sup ;i (h g) is a continuous mapping of Y pC > = g2T (equipped with the topology generated by the C ( )-one)>

(2:10:26)

(2:10:27)

concluding the proof of the lemma. Proof of Theorem 2.10.1. Let fhng1=1 be a minimizing sequence, i.e., n fhng1=1 Uad (2.10.28) n f (hn) ! h2U f (h): inf (2.10.29)
ad

C being the number from (2.10.9). Taking to notice (2.10.9), (2.10.28), and (2.10.31), we get ~ ^ h h h: (2:10:33) Combining (2.10.9), (2.10.28), (2.10.30), and Lemma 2.10.2 we obtain ~ lim i = 1 2 : : : q: i (h) = m!1 i (hm ) 0

By virtue of (2.10.9) and (2.10.10), the sequence fhng is bounded in Wp1 ( ), p > 2 let us chose a subsequence fhmg from it such that ~ hm ! h weakly in Wp1 ( ) (2.10.30) ~ hm ! h strongly in C ( ): (2.10.31) The relations (2.10.28) and (2.10.30) yield ~ C lim inf khmkWp1 ( ) khkWp1 ( ) (2:10:32) m!1

150

Chapter 2. Optimal control by coe cients in elliptic systems

~ This equality and (2.10.32), (2.10.33) imply h 2 Uad . Further, using (2.10.11), (2.10.29), and (2.10.30), we deduce ~ lim f (hm ) = f (h) = inf f (h):
m!1 h2Uad

~ Hence, the function h0 = h is a solution to the problem (2.10.12).

2.11 The general control problem

So far we have studied optimal control problems connected with the bilinear form ah (see (2.1.4)) whose coe cients aij were functions of control h 2 Wp1 ( ). Now, we proceed study problems in which not only coe cients but also operators of the bilinear form depend on the control moreover, we suppose that the right hand sides of the equations depend on the control. We consider optimal control problems under restrictions on solutions to the equations and on eigenvalues. We do not specify the space of controls in these problems, which allows us to apply the results obtained to general situations. Let, as before,

2.11.1 Bilinear form aq and the corresponding equation

be a bounded, Lipschitz domain in R2 , (x y) 2 ,

W=

s=1

W2ls ( )

where ls 1, s = 1 2 : : : , and let V be a closed subspace of W with the norm of the space W . Let also B1 be a Banach space, let Q1 be a closed set in B1 endowed with the topology induced by the B1 -strong one, and let mappings r ! Pi(r) be de ned such that

r ! Pi(r) is a continuous mapping of Q1 into L(W L2 ( )), (2:11:1) i = 1 2 ::: k L(W L2 ( )) being the space of linear continuous mappings of W into L2 ( ). We are given the family of the bilinear, continuous forms ar on the space V V depending on a parameter r running through the set Q1 de ned by
ZZ X k ; ; ar (u v) = a(r ) Pi(r)u Pj(r)v ij i j =1

dx dy u v 2 V:

(2:11:2)

Here, Pi(r) u is the image of an element u 2 V under the mapping

Pi(r) 2 L(W L2 ( ))

2.11. The general control problem

151

and

a(r ) 2 L1 ( ) ij
Assume that

a(r ) = a(r) ij ji

r 2 Q1 i j = 1 2 : : : k:

(2:11:3) (2.11.4) (2.11.5)

r ! a(r ) is a continuous mapping of Q1 into L1 ( ) (i j = ij 1 2 : : : k),

k k X (r ) X 2 aij (x y) i j cr i i j =1 i=1 (x y) 2 2 Rk cr = const > 0

9 > > > > = > > > >

where cr depends on r 2 Q1 .

Due to (2.11.3), the form ar is symmetric:

ar (u v) = ar (v u)
i.e.,

u v 2 V r 2 Q1 :

(2:11:6)

(2.11.1){(2.11.3) imply the form ar to be continuous on V V for any r 2 Q1 ,

jar (u v)j cr kukV kvkV ~ u v2V (2:11:7) cr being a positive number depending on r 2 Q1 . ~ Further, suppose that, for all r 2 Q1 , the system of operators fPi(r)g is
coercive in V , i.e.,
ZZ X; k Pi(r) u 2 dx dy i=1

r kuk2 V

u 2 V r = const > 0:

(2:11:8)

The relations (2.11.2), (2.11.5), and (2.11.8) yield the form ar to be coercive in V for all r 2 Q1 , i.e.,

ar (u u)

r kuk2 V

u 2 V r = const > 0:

(2:11:9)

We stress that the number r depends on r 2 Q1 . Now, consider the following problem: Given r 2 Q1 and f 2 V , nd a function ur f such that

ur f 2 V

ar (ur f v) = (f v)

v 2 V:

(2:11:10)

Theorem 2.11.1 Let ar be a bilinear form on V V de ned by the relations (2.11.1){(2.11.5), and let the inequality (2.11.8) holds. Then, for any r 2 Q1 and f 2 V , the problem (2.11.10) has a unique solution and the function r f ! ur f determined by this solution is a continuous mapping of Q1 V into V .

152

Chapter 2. Optimal control by coe cients in elliptic systems

Proof. The existence and uniqueness of a solution to the problem (2.11.10) is a consequence of the relations (2.11.6), (2.11.7), (2.11.9), and of the Riesz theorem. Let frn fn g1 n=1 Q1 V and rn ! r0 in Q1 (2.11.11) fn ! f0 in V : (2.11.12)
To prove the theorem, we need to establish that

urn fn ! ur0 f0 in V
where the functions urn fn satisfy the conditions

(2:11:13) (2:11:14) (2:11:15)

urn fn 2 V

arn (urn fn v) = (fn v)

v 2 V n = 0 1 2 :::
n kuk2 V

Let us show that

jarn (u u) ; ar0 (u u)j

lim = 0: n!1 n

u2V

Indeed, for every addend on the left hand side of (2.11.15), due to (2.11.2), we have for any u 2 V
ZZ h

a(rn) Pi(rn ) u Pj(rn ) u ; a(r0 ) Pi(r0 ) u Pj(r0 ) u dx dy ij ij

ZZ n ;

a(rn) Pi(rn ) u Pj(rn ) u ; Pj(r0 ) u + Pj(r0 ) u Pi(rn) u ; Pi(r0 ) u ij

; ; o

h;

c1 a(rn) L1( ) kuk2 Pi(rn) L(W L2 ( )) Pj(rn ) ; Pj(r0 ) L(W L2 ( )) V ij + Pi(rn) ; Pi(r0 ) L(W L2 ( )) + c2 kuk2 a(rn) ; a(r0 ) L1( ) u2V V ij ij
and (2.11.16) imply (2.11.15). For all r 2 Q1 , the bilinear form ar satis es the formula

+ a(rn) ; a(r0 ) Pi(r0 ) u Pj(r0 ) u ij ij

dx dy

(2.11.16)

c1 and c2 being positive numbers. Now, the relations (2.11.1), (2.11.4), (2.11.11), ar (u v) = (Ar u v) u v2V
(2:11:17)

where Ar 2 L(V V ), so that the problem (2.11.10) is equivalent to the following one ur f 2 V Ar ur f = f: (2:11:18)

2.11. The general control problem

153

By virtue of (2.11.6), the operator Ar is selfadjoint , i.e., (Ar u v) = (Ar v u) u v 2 V: Hence, for all r 2 Q1 , the following relation holds kAr kL(V V ) = sup j(Ar u u)j:
kukV =1

Denote by U the set of the invertible elements of the space L(V V ) equipped with the topology induced by that of L(V V ). Since for all r 2 Q1 the bilinear form ar is symmetric, continuous, and coercive on V V , the Riesz theorem together with (2.11.17) yields that Ar 2 U for all r 2 Q1 . From (2.11.11) and (2.11.19), we deduce r ! Ar to be a continuous mapping of Q1 into U . Now, taking into account (2.11.11), (2.11.12), (2.11.18) and applying Theorem 1.8.1, we get (2.11.13). The theorem is proved.

Thus, from (2.11.15), (2.11.17), we conclude nlim kArn ; Ar0 kL(V V ) = 0: !1

(2:11:19)

2.11.2 Bilinear form br and the spectral problem

Suppose that Y is a Banach space, V Y (L2 ( )) , the imbedding (2:11:20) V ! Y is compact. Suppose we are given a family of bilinear, symmetric, continuous forms br on Y Y depending on a parameter r running through Q1, i.e., br (u v) = br (v u) u v2Y (2.11.21) jbr (u v)j cr kukY kvkY u v 2 Y: (2.11.22) cr being a positive number depending on r 2 Q1 . The dependence of the bilinear form br on a parameter r 2 Q1 de nes the mapping r ! br which is supposed to satisfy the following condition r ! br is a continuous mapping of Q1 into L2 (Y Y R), g (2:11:23) L2 (Y Y R) being the vector normed space of bilinear continuous forms on Y Y equipped with the norm kbr kL2 (Y Y R) = sup jbr (u v)j: Given an element r 2 Q1 , consider the following eigenvalue problem
; (r) (r) ui i
kukY

1 kv kY 1

2V R

u(r) 6= 0 i

154

Chapter 2. Optimal control by coe cients in elliptic systems

(r) a ;u(r)

r i

; v = br u(r) v i

v 2 V:

(2.11.24)

By virtue of Theorem 1.5.8, the problem (2.11.24) has a countable set of eigenvalues f (ir)g1 , which are ordered so that (cf. subsec. 2.5.2) i=1

j
Q1 into `1 0 .

(r) j 1

(r) j 2

So, we have de ned the mapping r !

lim (r) = 0: i!1 i (r) = f (r) g1 of the topological space i i=1

Theorem 2.11.2 Let ar be a bilinear form on V V determined by the relations (2.11.1){(2.11.5) and (2.11.8). Let also (2.11.20) hold and let the bilinear form br satisfy the conditions (2.11.21){(2.11.23). Then, the function r ! (r) = f (ir)g1 , where (ir) are the eigenvalues of the problem (2.11.24), is a continuous i=1
mapping of Q1 into `1 0 :

Proof. Let frng1=1 be a sequence such that n frn g1=1 Q1 rn ! r0 in Q1 : n

arn ! ar0 in L2 (V V R):
By (2.11.23) and (2.11.25), we get

(2:11:25)

Then, the argument from the proof of Theorem 2.11.1 (see (2.11.15)), together with Remark 1.5.3 implies that (2:11:26) (2:11:27)

brn ! br0 in L2 (Y Y R):

i

Now, taking into account (2.11.26), (2.11.27) and applying Theorem 1.5.9, we obtain lim sup (rn) ; (r0 ) = 0 i i n!1 which concludes the proof of the theorem.

2.11.3 Basic control problem

Setting of the problem. Existence theorem

In subsection 2.11.1, we introduced the Banach space B1 and the closed set Q1 in B1 equipped with the topology induced by the B1 -strong one. Additionally, assume that B2 is a re exive Banach space, B2 B1 , and the imbedding (2.11.28) B2 ! B1 is compact, Q2 is a convex, closed, bounded set in B2 , Q2 Q1 , g (2.11.29)

2.11. The general control problem

155 (2.11.30) (2:11:31)

9

Q3 is an open set in B2 , Q2 Q3 Q1.

De ne a set of admissible controls by

Uad = r j r 2 Q2
Here

k (r)

0 k = 1 2 ::: l :

r ! k (r) is a continuous mapping of Q3 (endowed with= the topology generated by the B2 -weak one) into R, k = 1 2 : : : l.
We suppose that a goal functional ' satis es the condition

(2:11:32)

r ! '(r) is a continuous mapping of Q3 (equipped with the topology generated by the B2 -weak one) into R. r0 2 Uad '(r0 ) = r2U '(r): inf
ad

(2:11:33)

The optimal control problem consists in nding an element r0 such that (2:11:34)

Theorem 2.11.3 Let a nonempty set Uad be determined by the relations Proof. Let frng1=1 be a minimizing sequence, i.e., n frn g1=1 Uad inf n nlim '(rn ) = r2Uad '(r): !1
rm ! y weakly in B2 y 2 Q2 :

(2.11.28){(2.11.32), and let (2.11.33) hold. Then, the problem (2.11.34) has a solution. (2:11:35)

By virtue of (2.11.28) and (2.11.29) we can choose from the sequence frn g a subsequence frm g such that (2.11.36) (2.11.37)

Since rm 2 Uad , (2.11.32) and (2.11.36) yield that k (y) 0, k = 1 2 : : : l. Therefore, from (2.11.37) we get y 2 Uad. Finally, the relations (2.11.33), (2.11.35), and (2.11.36) imply that lim inf m!1 '(rm ) = '(y ) = r2Uad '(r): So, the function r0 = y is a solution to the problem (2.11.34).

156

Chapter 2. Optimal control by coe cients in elliptic systems

Approximate solution of the basic problem

In the same way as it was done before (see, for example, Section 2.3) we can search approximate solutions to the problem (2.11.34) on the set Hn \ Uad (fHn g being a sequence of nite-dimensional subspaces of B2 satisfying the limit density condition). If there exists a sequence of interior points of the set Uad convergent to the solution r0 of the problem (2.11.34) (in this case, the set Uad is equipped with the topology generated by that of the space B2 ), then the corresponding sequence of approximate solutions is a minimizing one and a subsequence that converges to r0 weakly in B2 may be extracted. Below we will consider another approach to approximate solution of the problem (2.11.34) that makes no use of the existence of a sequence of interior points from Uad convergent to r0 . Let a family of sets Q(t) depending on parameter t running through the set 2 1 ], > 1, be given such that

Q(t) is a convex, closed, bounded set in B2 for any t 2 1 ], (2.11.38) 2 ) Q2 Q2 (t) Q(t) Q3 for any t 2 (1 ], Q2 = Q(1) , (2.11.39) 2 2 Q(t1 ) Q2 (t2 ) if t2 > t1 . 2
Here Q2 (t) is the interior of the set Q(t), the latter being equipped with the topology 2 generated by the B2 -one. Additionally, we suppose that
t!1 q2Q(t) r2Q2 2

lim

sup inf kq ; rkB2 = 0:

(2:11:40)

The relation (2.11.40) means that the distance between an element q 2 Q(t) 2 and the set Q2 tends to zero uniformly in q 2 Q(t) as t ! 1. 2 For t 2 (1 ], de ne a set

U (t) = r j r 2 Q(t) 2

k (r)

t;1 k = 1 2 ::: l :

(2:11:41)

Here, k are the functionals from (2.11.31). The following statement is a direct consequence of the proof of Theorem 2.11.3.

Theorem 2.11.4 Let a set U (t) be de ned by the relations (2.11.38), (2.11.39), (2.11.41), and (2.11.32), and let the set U (1) = Uad be not empty. Assume that a function ' satis es the condition (2.11.33). Then, for all t 2 1 ], there exists an
element rt such that

rt 2 U (t)

'(rt ) = inf(t) '(r):

r 2U

(2:11:42)

2.11. The general control problem

157

Let fHn g1 be a sequence of nite-dimensional subspaces of B2 meeting the n=1 conditions lim inf kq ; rkB2 = 0 q 2 B2 (2.11.43) n!1 r2Hn Hn Hn+1 8n: (2.11.44) Consider the problem of nding a function rn such that rn 2 Hn \ U (t) (2.11.45) '(rn ) = inf (t) '(r): (2.11.46)

Theorem 2.11.5 Let the assumptions of Theorems 2.11.3 and 2.11.4 be ful lled, let fHng1 be a sequence of nite-dimensional subspaces of B2 satisfying the conn=1 ditions (2.11.43) and (2.11.44), and let r0 , rt be solutions to the problems (2.11.34) and (2.11.42), respectively, where t 2 (1 ]. Suppose that the set H1 \ U (t) is not
empty. Then, for any n, the problem (2.11.45), (2.11.46) has a solution rn , and From the sequence frn g one can choose a subsequence frm g such that

Hn \ U

'(rt ) nlim '(rn ) '(r0 ): !1

(2:11:47)

Proof. By hypothesis, the set H1 \U (t) is not empty. So, (2.11.44) implies Hn \U (t) to be not empty for all n. Since Hn \ U (t) is a compactum, (2.11.33) yields that

rm ! r weakly in B2 ~ r 2 U (t) ~ lim '(rm ) = nlim '(rn ) = '(~): r m!1 !1

(2.11.48)

problem (2.11.45), (2.11.46) has a solution for all n. By (2.11.42) and (2.11.46), '(rt ) '(rn ) 8n: (2:11:49) The relations (2.11.44) and (2.11.46) imply '(rn+1 ) '(rn ). Therefore, from (2.11.49) we deduce that the number sequence f'(rn )g1 converges and n=1 lim '(r ) '(rt ): (2:11:50) n!1 n By using (2.11.31), (2.11.32), (2.11.39), and (2.11.41), it is easy to see that r0 2 U (t) (U (t) being the interior of U (t) which is equipped with the topology generated by the B2 -one). This together with (2.11.43) implies the existence of a sequence fzng such that zn 2 Hn \ U (t) zn ! r0 in B2 : (2:11:51) Hence, lim '(z ) = '(r0 ): (2:11:52) n!1 n

158

Chapter 2. Optimal control by coe cients in elliptic systems

From (2.11.45), (2.11.46), and (2.11.51), we have '(zn ) '(rn ). Thus, (2.11.50) and (2.11.52) yield (2.11.47). By (2.11.38), we get that the sequence frn g is bounded in B2 . Choose a subsequence frm g from it such that

rm ! r ~

weakly in B2 :

(2:11:53)

Using (2.11.32), (2.11.38), and (2.11.53), we easily obtain r 2 U (t) . Finally, ~ (2.11.33) and (2.11.53) imply (2.11.48).

Remark 2.11.1 We did not use the supposition (2.11.40) in Theorem 2.11.5.
However, if (2.11.40) holds, then for t su ciently close to 1, \the getting of the functions rn and r out of the domain Uad can be made as small a desired, and in ~ this case (2.11.47) and (2.11.48) hold.

2.11.4 Application of the basic control problem (combined problem)

Suppose that

u v ! ar (u v) is a bilinear form on V V depending on= parameter r 2 Q1 and determined by the relations (2.11.1){
(2.11.5), (2.11.8), 9 u v ! br (u v) is a bilinear, symmetric, continuous form on= Y Y depending on parameter r running through Q1 and (2.11.20), (2.11.23) hold, r ! fr is a continuous mapping of Q1 into V :

(2.11.54) (2.11.55) (2.11.56)

Provided r 2 Q1 is given, consider the following problems. First problem: Find a function ur such that

ur 2 V
(r) a ;u(r)

ar (ur v) = (fr v)
; v = br u(r) v i

v 2 V: v2V

(2:11:57) (2:11:58)

Second problem: Find eigenvalues (r) of the equation i

i r i

A solution to the second problem determines a continuous mapping r ! (r) = f (r) g1 of the set Q1 (endowed with the topology generated by the i i=1 B1 -one) into `1 0 (see Theorem 2.11.2). We will control by the function r. In this connection, de ne a set of admissible controls as

u(r) being the eigenfunctions, u(r) 2 V . i i

Uad = r j r 2 Q2

(1) (r

u r ) 0 (k = 1 2 : : : n1 )

2.11. The general control problem

(2) (r

159

Theorem 2.11.6 Let the relations (2.11.54){(2.11.56) hold and let a nonempty set Uad be determined by the formulas (2.11.28){(2.11.30), (2.11.59){(2.11.61). Let also the goal functional meet the condition (2.11.33). Then, the problem (2.11.62) has a solution. Proof. To apply Theorem 2.11.3, de ne functionals k (r) in the following way: (1) k = 1 2 : : : n1 (2.11.63) k (r) = k (r ur ) (2) i = 1 2 : : : n2 : (2.11.64) n1 +i (r) = i (r (r)) The functionals k (r), k = 1 2 : : : n1 + n2 , are considered on the set Q3 (see (2.11.30)). Further, let frn g1 n=1 Q3 , r0 2 Q3 , and rn ! r0 weakly in B2 . Then, by virtue of (2.11.28), rn ! r0 strongly in B1 , i.e., rn ! r0 in Q1 (2:11:65) and taking into account (2.11.56), we deduce frn ! fr0 in V . Now, Theorem 2.11.1 yields urn ! ur0 in V: (2:11:66) By using (2.11.60), (2.11.65), and (2.11.66), we get lim (1)(r u ) = (1) (r0 ur0 ) k = 1 2 : : : n1 : k n!1 k n rn
So, the functionals k (r) determined by the formula (2.11.63) satisfy the condition (2.11.32). By (2.11.65) and Theorem 2.11.2, we obtain (rn ) ! (r0 ) in `1 0 . Therefore, from (2.11.61), (2.11.65) we deduce that lim (2) (r (rn )) = (2) (r0 (r0 )) i = 1 2 : : : n2 i n!1 i n

(r)) 0 (i = 1 2 : : : n2 ) :(2.11.59) Here, the set Q2 is determined by the relations (2.11.28), (2.11.29), and ur , (r) are solutions to the rst and second problems, r w ! (1)(r w) is a continuous mapping of Q1 V into (2.11.60) k R (k = 1 2 : : : n1 ), r z ! (2)(r z ) is a continuous mapping of Q1 `1 0 into (2.11.61) i R (i = 1 2 : : : n2 ). Provided the goal functional '(r) satis es the condition (2.11.33), consider the optimal control problem: Find a function r0 such that r0 2 Uad '(r0 ) = r2U '(r): inf (2:11:62)
i
ad

i.e., the functionals k (r) determined by the formula (2.11.64) meet the condition (2.11.32). Now, Theorem 2.11.3 yields that the problem (2.11.62) has a solution. Notice that in order to solve the problem (2.11.62) approximately, one can use the technique from subsec. 2.11.3.

160

Chapter 2. Optimal control by coe cients in elliptic systems

2.12 Optimization by the shape of domain and by operators

Hitherto, the domain was assumed to be xed. However, in applications, there arise many problems connected with the search of an optimal domain. So, we proceed now study problems in which controls are the shape of domain and operators. Let be a bounded region in R2 , let Ml be a topological space (l being a natural number), and let to any q 2 Ml there correspond a domain q in R2 and a di eomorphism Pq of the set q onto such that ; ; Pq 2 C l q Pq;1 2 C l (2:12:1) q where Pq;1 is the ; inverse mapping of Pq . The relations (2.12.1) mean that Pq = ; (Pq; Pq2 ), Pq;1 = ; q;1 Pq;1 , where Pqi and Pqi 1 are scalar functions and Pqi 2 P1 2 1 ; C l q , Pqi 1 2 C l , i =Q 2. 1 Q Further, let W ( ) = i=1 W2li ( ), W ( q ) = i=1 W2li ( q ), and let V ( ), V ( q ) be closed subspaces of W ( ) and W ( q ), respectively. Suppose that l = maxi li . Then, from (2.12.1) and Theorem 1.14.3, it follows that the mapping u ! u Pq is an isomorphism of W ( ) onto W ( q ). Assume the following condition to be ful lled: for all q 2 Ml the mapping u ! u Pq is an isomorphism (2:12:2) of V ( ) onto V ( q ). Provided Y is a Hilbert space, we will refer to the set of the bilinear, symmetric, continuous, coercive forms on Y Y as L2 sc (Y Y R). Suppose we are given a form aq 2 L2 sc (V ( q ) V ( q ) R): Denote by Pq aq the bilinear form on V ( ) V ( ) that is the image of the form aq under the mapping Pq and is given by Pq aq (u v) = aq (u Pq v Pq ) u v 2 V ( ): (2:12:3) To get the explicit form of Pq aq , we have to change the variable y 2 q by that x = Pq (y) 2 . We stress that Pq aq 2 L2 sc (V ( ) V ( ) R). Indeed, the form Pq aq is symmetric and continuous because of (2.12.2), (2.12.3), and of the symmetricity and continuity of the form aq . Since aq is coercive, by (2.12.2) and (2.12.3), we have Pq aq (u u) = aq (u Pq u Pq ) cku Pq k2 ( q ) c1 kuk2 ( ) V V u 2 V ( ) c1 = const > 0:

2.12.1 Domains and bilinear forms

2.12. Optimization by the shape of domain and by operators

161

2.12.2 Optimization problem connected with solution of an operator equation

Let G be a topological space such that from every sequence fhn g G one can choose a subsequence fhm g convergent to some h 2 G. (2:12:4) (2:12:5)

Let also to every pair (h q) 2 G Ml there correspond a form ahq such that

ahq 2 L2 sc (V ( q ) V ( q ) R): Pq ahq 2 L2 sc (V ( ) V ( ) R):

Suppose that

De ne a form Pq ahq by the equality (2.12.3) in which aq is replaced by ahq . By the above argument, we have

h q ! Pq ahq is a continuous mapping of G Ml into L2 sc (V ( ) V ( ) R).

(2:12:6)

Assume we are given a mapping B satisfying the following condition h q ! B (h q) is a continuous mapping of G Ml into (2:12:7) (V ( )) . Now, consider the problem: Given an element (h q) 2 G Ml , nd a function uhq such that ; uhq 2 V ( q ) ahq (uhq v) = B (h q) v Pq;1 v 2 V ( q ): (2:12:8) Since B (h q) 2 (V ( )) , the relation (2.12.2) yields that ; v ! B (h q) v Pq;1 is a linear continuous mapping of V ( q ) into R. This together with (2.12.5) implies

that the problem (2.12.8) has a unique solution. By virtue of (2.12.2) and (2.12.3), the problem (2.12.8) is equivalent to the following one: Find a function uhq such that ~

uhq 2 V ( ) ~ Pq ahq (~hq w) = (B (h q) w) u w 2V( ) and uhq = uhq Pq;1 . ~ Let a topological space Nl be de ned such that 9 from every sequence fqn g Nl one can choose a subse-= quence fqm g convergent to some q 2 Nl , Nl Ml , the imbedding Nl into Ml being continuous.

(2:12:9)

(2:12:10)

162

Chapter 2. Optimal control by coe cients in elliptic systems

Assume that functionals i be de ned such that h q u ! i (h q u) is a continuous mapping of G Nl (2:12:11) V ( ) into R (i = 0 1 2 : : : m). De ne a set of admissible controls as Uad = (h q) j (h q) 2 G Nl i (h q uhq ) 0 i = 1 2 : : : m ~ (2:12:12) and let a goal functional be of the form J (h q) = 0 (h q uhq ) ~ (2:12:13) uhq being a solution to the problem (2.12.9). ~ The optimal control problem consists in nding an element (h0 q0 ) such that (h0 q0 ) 2 Uad J (h0 q0 ) = inf J (h q): (2:12:14)

Theorem 2.12.1 Let the conditions (2.12.2){(2.12.7) and (2.12.9){(2.12.13)

hold, and let the set Uad be nonempty. Then, the problem (2.12.14) has a solution.

(h q)2Uad

Proof. Let fhn qng be a minimizing sequence, i.e., fhn qn g Uad lim J (hn qn ) = (h qinfU J (h q): n!1 )2 ad

(2:12:15)

(2.12.4) and (2.12.10) imply that we can choose a subsequence fhm gmg from the sequence fhn qn g such that hm ! h0 in G qm ! q0 in Nl qm ! q0 in Ml : (2:12:16) From here, taking into account (2.12.6) and (2.12.7), we get Pqm ahm qm ! Pq0 ah0 q0 in L2 sc (V ( ) V ( ) R) B (hm qm ) ! B (h0 q0 ) in (V ( )) : (2.12.17) Using Theorem 1.8.1 and (2.12.17), we obtain uhmqm ! uh0 q0 in V ( ) ~ ~ (2:12:18) where uhm qm and uh0 q0 are solutions to the problem (2.12.9) for h = hm , q = qm ~ ~ and h = h0 , q = q0 , respectively. Now, taking into consideration (2.12.11), (2.12.12), (2.12.16), and (2.12.18) we deduce that (h0 q0 ) 2 Uad. Further, due to (2.12.11), (2.12.13), (2.12.15), (2.12.16), and (2.12.18), we conclude J (h0 q0 ) = inf J (h q):
(h q)2Uad

Thus, the pair h0 = h0 q0 = q0 is a solution to the problem (2.12.14).

2.12. Optimization by the shape of domain and by operators

163

2.12.3 Eigenvalue optimization problem

V( )

Under the conditions imposed above, we suppose also that Y ( ) and Y ( q ) are Banach spaces of functions de ned on and q , respectively, such that

Y ( ), the imbedding of V ( ) into Y ( ) is com- (2.12.19) pact, for all q 2 Ml , the mapping u ! u Pq is an isomorphism (2.12.20) of Y ( ) into Y ( q ).
By (2.12.2), (2.12.19), and (2.12.20), we get that

V ( q ) Y ( q ), the imbedding of V ( q ) into Y ( q ) is compact for all q 2 Ml .

(2:12:21)

Further, let to every pair (h q) 2 G Ml there corresponds a bilinear, symmetric, continuous form bhq on Y ( q ) Y ( q ). We denote by Pq bhq the bilinear, symmetric, continuous form on Y ( ) Y ( ) that is determined by

Pq bhq (u v) = bhq (u Pq v Pq )
Assume that

u v 2 Y ( ):

(2:12:22)

h q ! Pq bhq is a continuous mapping of G Ml into (2:12:23) L2 (Y ( ) Y ( ) R). Now, consider the problem: Given a pair (h q) 2 G Ml , nd the eigenvalues i f (hq) g of the following problem: u(i) 2 V ( q ) u(i) 6= 0 hq ; hq (i) a u(i) v = b ;u(i) v hq hq hq hq hq
(i)

v 2 V ( q ):

hq

2R

(2.12.24)

Making use of (2.12.5), (2.12.21), and of Theorem 1.5.8, we deduce that there exists a sequence of the eigenfunctions of the problem (2.12.24) convergent to zero, i.e., f (i) g1 = hq 2 `1 0 and j (1) j j (2) j j (3) j . By virtue of (2.12.2), hq i=1 hq hq hq (i) are the eigenvalues of the following problem (2.12.3), and (2.12.22), hq

u(i) 2 V ( ) ~hq u(i) 6= 0 ~hq (i) P a ;u(i) w = P b ;u(i) w q hq ~hq hq q hq ~hq

(i)

hq

2R w 2V( )

(2.12.25)

where u(i) = u(i) Pq;1 , u(i) being the eigenfunctions of the problem (2.12.24). ~hq hq hq Further, suppose we are given mappings Ai and j such that

! Ai is a continuous mapping of `1 0 into R (i = (2.12.26) 1 2 : : : k),

164

Chapter 2. Optimal control by coe cients in elliptic systems

h q ! j (h q) is a continuous mapping of G Nl into R (2.12.27) (j = 0 1 2 : : : m).

De ne a set of admissible controls as

Uad = (h q) j (h q) 2 G Nl j (h q) 0 (j = 1 2 : : : m) Ai hq 0 (i = 1 2 : : : k) : (2.12.28)
Consider the problem: Find a pair (h0 q0 ) such that (h0 q0 ) 2 Uad
0 (h0 q0 ) = (h qinfU )2
ad

0 (h

q):

(2:12:29)

Theorem 2.12.2 Let the conditions (2.12.2){(2.12.6), (2.12.10), (2.12.19),

(2.12.20), (2.12.22), (2.12.23), and (2.12.26){(2.12.28) hold, and let the set Uad be nonempty. Then, the problem (2.12.29) has a solution.

Proof. By using (2.12.6), (2.12.23), (2.12.25), and Theorem 1.5.9, we conclude that h q ! hq is a continuous mapping of G Ml into `1 0. Now, choosing from the minimizing sequence a convergent subsequence and passing to the limit, we see that the limit element is a solution of the problem (2.12.29). Remark 2.12.1 In some important problems, the spaces V ( q ) are de ned so that the condition (2.12.2) isQnot satis ed, while the mapping u ! u Pq is Q
an isomorphism of W ( ) = i=1 W2li ( ) onto W ( q ) = i=1 W2li ( q ) for an arbitrary q 2 Ml , l = maxi li . Particularly, (2.12.2) does not hold when V ( q ) consists of solenoidal vector-valued functions. In this case, one can reformulate the direct problem. The general approach to such optimization problems is considered in Litvinov (1989).

2.12.4 Some realizations of the spaces Ml and Nl

Double-connected domain
Suppose Q is a double-connected domain in R2 that is star-shaped with respect to the origin (Fig. 2.12.1), i.e., the boundary S of the domain Q consists of two connected components S1 and S2 . Suppose also that S1 is given by the equations

y1 = q(1) ( ) cos
and the equations for S2 are of the form

y2 = q(1) ( ) sin y2 = q(2) ( ) sin :

(2:12:30) (2:12:31)

y1 = q(2) ( ) cos

Here, 0 2 and the functions q(i) characterizing the domain Q are periodic, (i) (0) = q (i) (2 ), i = 1 2. i.e., q

2.12. Optimization by the shape of domain and by operators

y2 S2

165

Q=W
0
a
S1

y1

Figure 2.12.1: Double-connected star-shaped domain Q Assume that

r1 r2 " are positive numbers r1 + " < r2 : (2:12:32) De ne a set Ml in the following way: ~ Ml = q = (q(1) q(2) ) j q 2 (C l ( 0 2 ]))2 r1 q(1) ( ) (1) ( ) + " q (2) ( ) r q 2 2 0 2 ] : (2.12.33)
~ Here, C l ( 0 2 ]) is the space of functions de ned on the interval 0 2 ] having continuous derivatives up to the l-th order and satisfying the periodicity condition, i.e.,
n o dk dk ~ C l ( 0 2 ]) = f f 2 C l ( 0 2 ]) d f (0) = d f (2 ) k = 0 1 2 : : : l : k k (2:12:34) ~ The space C l ( 0 2 ]) can be considered as the restriction on the interval

0 2 ] of the set of 2 -periodical functions de ned on R and having continuous derivatives in R up to the l-th order. De ne a set Nl by ~ Nl = q = (q(1) q(2) ) q 2 W2l+1 (0 2 )
;
2

r1 q(1) ( ) q(1) ( ) + " q(2) ( ) r2

kqk;W l+1 (0 2 ~
2

r3

2 0 2 ] : (2.12.35)

~ ~ Here, W2l+1 (0 2 ) is the closure of C l+1 ( 0 2 ]) in W2l+1 (0 2 ), and r3 is a positive number. Now, if we endow Ml;and Nl with the strong topology of the space C l ( 0 2 ])2 ~ and the weak topology of W2l+1 (0 2 ) 2 , respectively, then, due to the imbedding

166

Chapter 2. Optimal control by coe cients in elliptic systems

theorem, the condition (2.12.10) is satis ed. Moreover, to every q 2 Ml there corresponds the double-connected domain q R2 the connected components of the boundary of which are de ned by the formulas (2.12.30) and (2.12.31). De ne a set by = x = (x1 x2 ) j x 2 R2 1 < x2 + x2 < 4 : 1 2 (2:12:36) Let F stand for the mapping transferring polar coordinates of a point into Cartesian ones:

F : (r ') ! F (r ') = (x1 x2 )

x1 = r cos ' x2 = r sin '

(2:12:37)

F is a bijection of the set Q1 = (r ') j (r ') 2 R2 0 < r < 1 0 ' < 2 onto the complement of the origin in R2 , and the restriction of F on Q2 = (r ') j 0 < r < 1 0 < ' < 2 g is a di eomorphism of the C 1 class of the set Q2 onto F (Q2 ). Denote by F ;1 the mapping inverse of F . Then, for q 2 Ml , de ne a mapping Pq : q ! as Pq = F D F ;1 (2:12:38) where D : F ;1 ( q ) ! F ;1 ( ), ( ) ! D( ) = (r ') ; 2q(1) ( ) + q(2) ( ) r = q(2) ( ) ; q(1) ( ) '= (2.12.39)

by the relations (2.12.37){(2.12.39) to satisfy the condition (2.12.1).

D is a bijection, and the inverse bijection D;1 : F ;1 ( ) ! F ;1 ( q ) is of the form (r ') ! D;1 (r ') = ( ) (1) (') ; q (2) (') + q (2) (') ; q (1) (') r = 2q = ': (2.12.40) Taking to notice that q 2 Ml , we easily conclude the mapping Pq determined

Single-connected domain

Now, de ne sets Ml and Nl by the following relations ~ Ml = q j q 2 C l ( 0 2 ]) r1 q( ) r2 2 0 2 ] l+1 (0 2 ) kq k ~ l+1 ~ Nl = q j q 2 W2 W2 (0 2 ) r3 r1 q ( ) r2

202 ]

(2.12.41)

2.12. Optimization by the shape of domain and by operators

167

where r3 = const > 0 and r1 , r2 meet (2.12.32). We endow Ml with the strong ~ topology of the space C l ( 0 2 ]) and Nl by the weak topology of W2l+1 (0 2 ). Then, due to the imbedding theorem, the condition (2.12.10) holds. To every q 2 Ml there corresponds the single-connected, star-shaped domain q R2 whose boundary is de ned by the function ! q( ), where is the polar angle. Denote by B (0 c) the open ball in R2 centered at 0, of radius c > 0, i.e., B (0 c) = f(x1 x2 ) j 0 x2 + x2 < c2 g, and let 1 2 = B (0 r1 ) r4 2 R 0 < r4 < r1 : (2:12:42) De ne a mapping Pq;1 : ! q by if (x1 x2 ) 2 B (0 r4 ) Pq (x1 x2 ) = (x1 x2 ) ;1 F D1 F (x1 x2 ) if (x1 x2 ) 2 n B (0 r4 ):
;1

(2.12.43) (2.12.44)

Here,

D1 : F ;1 ( n B (0 r4 )) ! D1 (F ;1 ( n B (0 r4 )))
r ' ! D1 (r ') = (
) (r ') =
l+1 X i=0

ai (')ri

= ' (2.12.45)

and the functions ' ! ai (') are determined by the relations (r4 ') = r4 (r1 ') = q(') @ (r ') = 1 @ k (r ') = 0 k = 2 3 : : : l: @r 4 @rk 4

(2.12.46)

Let us consider the case when l = 1. Then, the solution to the system (2.12.45), (2.12.46) is given by the formulas

We have @ = 1 + 2(r ; r4 ) a2 , and since q 2 M1 , it follows that a2 (') 0 @r for all ' 2 0 2 ]. Hence, @ 1 as r r4 , which implies that the mapping D1 @r is invertible. Therefore, there exists a mapping Pq inverse of Pq;1 . It is easy to see that Pq;1 2 C 1 ( q ), and the Jacobian of the mapping Pq;1 at every point (x1 x2 ) 2 is not equal to zero. Now, the results on the di erentiation of an implicit function (see Schwartz (1967)) imply that Pq 2 C 1 ( q ): Consider the case l = 2. Then, the solution to the system (2.12.45), (2.12.46) is given by the formulas

;r a2 = (rq ; r1)2
1 4

a1 = 1 ; 2r4 a2

2 a0 = q ; r1 a1 ; r1 a2 :

(2:12:47)

;r a3 = (rq ; r1)3
1 4

a2 = ;3r4 a3

2 a1 = 1 ; 2r4 a2 ; 3r4 a3

168

Chapter 2. Optimal control by coe cients in elliptic systems

2 3 a0 = q ; r1 a1 ; r1 a2 ; r1 a3 :

(2.12.48)

By (2.12.45) and (2.12.48), we conclude

1 as r r4 . Now, the argument similar to the above one shows that there exists a mapping Pq inverse of Pq;1 and

@ = 1 + 3a (r ; r )2 : 3 4 @r Since q 2 M2 , we get a3 (') 0 for all ' 2 0 2 ]. Hence, @ @r Pq;1 2 C 2 (

q)

Pq 2 C 2 ( q ):

Notice that the above stated approach to the domain shape optimization is based on the construction of di eomorphisms Pq of sets q , q 2 Ml , onto a xed set and on the transition to the optimization problem in the xed domain . In somewhat other form, this approach was suggested and analyzed by Murat and Simon (1976).

2.13 Optimization problems with smooth solutions of state equations

We have already studied optimization problems in which the weak (generalized) solutions of the state equations in the form of systems of elliptic equations were used. In some applications, though, smooth solutions of the state equations are only allowed. Particularly, such situation occurs in the optimization problems in which restrictions are imposed on the values of the solution of the state equation and on the derivatives of this solution at points of the domain under consideration. For example, in some formulations of optimization problems for elastic solids, plates, and shells with restrictions on sti ness and strength, the displacements and stresses must be bounded at each point of the body, the stresses being de ned via the derivatives of the function of displacements. Below, we consider domain shape optimization problems for objects described by systems of equations that are elliptic in the sense of Douglis{Nirenberg, see Agmon et al. (1964).

2.13.1 Systems of elliptic equations

A(x D)u(x) = f (x) B (x D)u(x) = g(x)

Let the state of some object be described by the following system of di erential equations:

x2 x 2 S:

(2.13.1) (2.13.2)

2.13. Optimization problems with smooth solutions of state equations

169

Here, is a bounded domain in Rn with a boundary S , u = (u1 : : : um) and f = (f1 : : : fm ) are m-dimensional vector-valued functions de ned in , x = (x1 : : : xn ) are points of , A(x D) is a square m m matrix with elements @ Aij (x D), D = (D1 : : : Dn ), Di = @xi , Aij are polynomials in D with coe cients depending on x 2 . The matrix A(x D) is supposed to satisfy the following conditions. The orders of the operators Aij (x D) depend on two systems of integers s1 : : : sm , max si = i 0, and t1 : : : tm , si corresponding to the i-th equation and tj to the j -th dependent variable uj , the order of the operator Aij (x D) does not exceed si + tj , and if si + tj < 0, then Aij (x D) = 0.

One supposes also that the following conditions hold: Condition of Ellipticity. At each point x 2 S , for an arbitrary = ( 1 : : : n ) 2 Rn , 6= 0, ; H(x ) = det A0ij (x ) 6= 0 where A0ij (x ) consists of the terms of Aij (x ) that are exactly of the order si + tj . The matrix with the elements A0ij (x ) is denoted by A0 (x ). Supplementary Condition. Let x 2 S , = + , where and are arbitrary tangent and normal vectors to S at x. Then, the polynomial H(x + ) in the complex variable has exactly r roots with positive imaginary part and r roots with negative imaginary part. Thus, the degree of the polynomial H(x ) (with respect to ) is 2r, and m X (si + ti ) = 2r: In (2.13.2), B (x D) is a rectangular matrix that has r rows and m columns with elements Bqj (x D), g = (g1 : : : gr ) is an r-dimensional vector-valued function de ned on S , Bqj (x D) are polynomials in D with coe cients depending on x 2 S . Let qj be the order of Bqj (x ), and let q = j=1 ::: m( qj ;tj ). Then, q +tj qj . max The sum of all the terms of the polynomial Bqj (x ) which are of degree q + tj 0 0 is denoted by Bqj (x ), and the matrix with the elements Bqj (x ) is designated 0 (x ). by B At any point x 2 S , for arbitrary tangent 6= 0 and normal to S , we denote by j+ (x ), j = 1 : : : r, the roots (in ) of the equation H(x + ) = 0 with positive imaginary part. The existence of these roots is assured by the supplementary condition. Set
i=1

M +(x

)=

r Y j =1

( ; j+ (x )):

^ Let A(x ) denote the matrix adjoint of A0 (x ): ^ A0 (x )A(x ) = H(x )I

170

Chapter 2. Optimal control by coe cients in elliptic systems

where I is the unit matrix. Complementing Boundary Condition. For any x 2 S and any nonzero vector tangent to S at x, the rows of the matrix ^ B 0 (x + )A(x + ) (which are polynomials in ) are linearly independent modulo M + (x ) (for a detailed description, see Agmon et al. (1964), Solonnikov (1964)). The problem (2.13.1), (2.13.2) is considered in the Holder spaces C l ( ), where l > 0 is not an integer. The C l ( ) is provided with the norm

kukC l( ) = kukC l]( ) +

k k 0 sup jD u(x) ;0D; u](x )j jx ; x jl l x x0 2 jkj= l]

X

(2.13.3)

where l] is an integer such that l ; l] 2 (0 1), jxj =

kukC l]( ) =

X
jk j

n X 2 1 xi 2 , and i=1

l] x2

max jDk u(x)j:

(2.13.4)

De ne spaces Vl and Hl as follows:

m Y l+t C j( ) j =1 m Y Hl = C l;si ( ) i=1

Vl =

r Y l; C q (S ): q=1

(2.13.5)

We also de ne an operator L 2 L(Vl Hl ) by

L : u ! Lu = (A(x D)u B (x D)u)

^ and let Vl = ker L, Hl = L(Vl ). Then, the following representations are valid: ^ ^ Vl = Vl Vl Hl = Hl Hl (2.13.6) ^ ^ where stands for direct sum of subspaces. The dimensions of Vl and Hl are ^l and Hl do not depend on l provided the coe cients of the operators ^ nite, and V A(x D), B (x D) and the boundary S are of the C 1 class. We say that (2.13.1), (2.13.2) is an elliptic problem if the condition of ellipticity, the supplementary condition, and the complementing boundary condition hold.

2.13. Optimization problems with smooth solutions of state equations

171

Theorem 2.13.1 Let (2.13.1), (2.13.2) be an elliptic problem, and let l be not an integer, l > max(0 1 : : : r ). Let also the boundary S be of the C l+tmax class, where tmax = max(t1 : : : tm ), and let the coe cients of the operators Aij (x D) and Bqj (x D) belong to C l;si ( ) and C l; q (S ), respectively. Then, the operator L is an isomorphism of Vl onto Hl : A proof of this theorem can be found in Solonnikov (1966). By using Theorem 2.13.1, it is easy to obtain the following Theorem 2.13.2 Assume the conditions of Theorem 2.13.1 hold and the dimen^ ^ ^ sions of Vl and Hl are equal. Let f'i gk=1 be a basis in Vl and let f i gk=1 be a basis i i ^ l . De ne an operator T 2 L(Vl Hl ) by in H
T'i = i i = 1 ::: k Tu = 0 u 2 Vl : Then, the operator L1 : u ! L1 u = Lu + Tu is an isomorphism of Vl onto Hl :

2.13.2 Elliptic problems in domains and in a xed domain

Let M be a space of controls. We assume that M is an open set in an a ne normed space X and M is provided with the topology generated by that of X . Suppose a domain q in Rn with a boundary Sq of the C l+tmax and a di eomorphism Pq of q onto of the C l]+1+tmax class are given for every q 2 M , that is,

Pq 2 C l]+1+tmax ( q ) Pq;1 2 C l]+1+tmax ( q ): (2.13.7) Then, the mapping u ! u Pq is an isomorphism of C p ( ) onto C p ( q ) and of C p (S ) onto C p (Sq ) for any p 2 0 l + tmax]. An elliptic problem in q of the above type is also given for every q 2 M Aq (y D)u(y) = fq (y) y2 q (2.13.8) Bq (y D)u(y) = gq (y) y 2 Sq : We denote by Vlq and Hlq the spaces Vl and Hl (see (2.13.5)) in which and S are substituted by q and Sq , respectively. We suppose also that 9 for any q 2 M , the operator > = Lq : u ! Lq u = (Aq (y D)u Bq (y D)u)> (2.13.9) is an isomorphism of Vlq onto Hlq .
e e e For every q 2 M , de ne an operator Lq = (Aq Bq ) 2 L(Vl Hl ) by e e e Lq u = (Aq u Bq u) e Aq u = (Aq (u Pq )) Pq;1 e Bq u = (Bq (u Pq )) Pq;1 :

(2.13.10)

172

Chapter 2. Optimal control by coe cients in elliptic systems

Here, Aq and Bq are the operators from (2.13.8), and we write Aq , Bq instead e of Aq (y D), Bq (y D). The operator Lq is obtained from the operator Lq under the change of variables corresponding to the mapping Pq . Owing to (2.13.7) and (2.13.9), we have
e the operator Lq is an isomorphism of Vl onto Hl , and if a function u 2 Vl is a solution of the problem ~

(2.13.11)

is a solution of (2.13.12). Suppose that

e~ Aq u = fq Pq;1 in (2.13.12) e ~ Bq u = gq Pq;1 on S where (fq gq ) 2 Hlq , then the function u = u Pq is a solution of the problem ~ (2.13.8). On the contrary, if u is a solution of the problem (2.13.8), then u = u Pq;1 ~ e e e q ! Lq = (Aq Bq ) is a continuous mapping of M into L(Vl Hl ), q ! (fq Pq;1 gq Pq;1 ) is a continuous mapping of M into Hl . )

(2.13.13) (2.13.14)

Theorem 2.13.3 Let the conditions (2.13.7), (2.13.9), (2.13.13), and (2.13.14) hold. Then, for every q 2 M , there exists a unique solution u of the problem ~ (2.13.12), and the function : q ! (q) = u determined by this solution is a ~

continuous mapping of M into Vl : Proof. The existence and uniqueness of the solution to the problem (2.13.12) follow from (2.13.7) and (2.13.9). Let q 2 M , fqk g1 M , and qk ! q in M . By k=1 (2.13.13) and (2.13.14), we have
e e Lqk ! Lq in L(Vl Hl ) (fqk Pq;1 gqk Pq;1 ) ! (fq Pq;1 gq Pq;1 ) k k

in Hl :

(2.13.15) (2.13.16)

e e From (2.13.15) and from the invertibility of the operators Lqk and Lq , we conclude the convergence of the inverse operators (e.g., Schwartz (1967)), that is,

in L(Hl Vl ): (2.13.17) Now, from (2.13.16) and (2.13.17), we obtain that (qk ) ! (q) in Vl , concluding the proof. Determine a mapping T : M Vl ! Hl by

eq eq L;k1 ! L;1

q 2 M u 2 Vl e e T (q u) = (Aq u ; fq Pq;1 Bq u ; gq Pq;1 ):

(2.13.18)

2.13. Optimization problems with smooth solutions of state equations

173

Obviously, the function : M ! Vl introduced in the formulation of Theorem 2.13.3 is the implicit function determined by the mapping T , that is,

T (q (q)) = 0

(2.13.19)

and (q) = u, where u is the solution of the problem (2.13.12). The existence and ~ ~ continuity of the implicit function follow from Theorem 2.13.3. Consider now the question on the di erentiability of .
e Theorem 2.13.4 Let the conditions (2.13.7), (2.13.9) hold and let q ! Lq = e e (Aq Bq ), q ! (fq Pq;1 gq Pq;1) be Frechet continuously di erentiable mappings of M into L(Vl Hl ) and into Hl , respectively. Then, the function determined by the equation (2.13.19) is a Frechet continuously di erentiable mapping of M into Vl , and the Frechet derivative 0 of at point q 2 M is given by 0 (q )h = ;L;1 @ T e (q (q))h h2X (2.13.20)

@q

where the operator @ T (q u) is given by the formula @q

@ T (q u)h = (A0 h)u ; (f P ;1 )0 h (B 0 h)u ; (g P ;1 )0 h eq eq q q q q @q h2X

(2.13.21)

Proof. It is obvious that T is a Frechet continuously di erentiable mapping of M Vl into Hl . Then, @ T (q u) is determined by the formula (2.13.21), and @q @ T (q u) = Lq = (Aq Bq ). From here, taking into account (2.13.9), we obtain that e e e @u the operator @ T (q u) is an isomorphism of Vl onto Hl . Now, the theorem follows @u from Theorem 1.9.2. Remark 2.13.1 We assumed above the condition (2.13.9) to hold. Let it be not satis ed. Then, upon (2.13.6), the following representations are valid ^ ^ Vlq = Vlq Vlq Hlq = Hlq Hlq (2.13.22)
^ where Vlq = ker Lq , Hlq = Lq (Vlq ). Suppose that ^ ^ dim Vlq = dim Hlq = kq q2M (2.13.23)

e e0 Here, A0q , Bq , (fq Pq;1 )0 , (gq Pq;1 )0 are the Frechet derivatives at a point q of e e the mappings q ! Aq , q ! Bq , q ! (fq Pq;1 ), q ! (gq Pq;1 ).

where kq is a natural number. Analogously to the operator T from Theorem 2.13.2, we de ne an operator Tq by

Tq 'qi = qi

i = 1 : : : kq

Tq u = 0

u 2 Vlq

(2.13.24)

174

Chapter 2. Optimal control by coe cients in elliptic systems

^ ^ where 'qi and qi are basis functions in Vlq and Hlq . Then, the operator Lq1 = Lq + Tq is an isomorphism of Vlq onto Hlq . Therefore, under the suitable conditions, the results stated above remain true if we substitute the operator Lq for Lq1 . It should be noted that ^ ^ {(Lq ) = dim Vlq ; dim Hlq is called the index of the elliptic operator Lq = (Aq Bq ). So, (2.13.23) means that {(Lq ) = 0 q 2 M: (2.13.25) It is known that the index is an unstable characteristic of an operator (see, e.g., Rempel and Schulze (1982)), and for an arbitrary elliptic operator Lq the condition ^ (2.13.25) may be unsatis able. But usually, in physical problems, the spaces Vlq ^ lq have some physical meaning that does not depend on a control q, i.e., and H on the shape of the domain q , and because of this meaning the dimensions of ^ ^ the spaces Vlq and Hlq are equal (see, e.g., Section 5.9). So, (2.13.23) is a natural assumption for applied problems.

2.13.3 The problem of domain shape optimization

Let functionals i over M Vl be given such that (q u) ! i (q u) is a continuous mapping of M Vl into (2.13.26) R, i = 0 1 : : : k. De ne functionals i over M by i = 0 1 ::: k (2.13.27) i (q ) = i (q (q )) where (q) is determined by the expression (2.13.19). Let M1 be a compact set in M . We take a set of admissible controls Uad in the form Uad = q j q 2 M1 i (q) 0 i = 1 2 : : : k : (2.13.28) The optimization problem consists in nding q0 such that q0 2 Uad inf 0 (q): (2.13.29) 0 (q0 ) = q2U

Theorem 2.13.5 Let the conditions (2.13.7), (2.13.9), (2.13.13), (2.13.14), and

ad

(2.13.26) hold, let M1 be a compact set in M , and let a nonempty set Uad be determined by (2.13.28). Then, there exists a solution of the problem (2.13.29).

that

Proof. As M1 is a nonempty set, there exists a minimizing sequence fqng such

fqn g Uad
lim 0 (qn ) = q2U inf
ad

0 (q ):

(2.13.30)

2.13. Optimization problems with smooth solutions of state equations

175

As M1 is a compact set in M , we can choose a subsequence fqm g such that qm ! z in M , z 2 M1 . By Theorem 2.13.3, (qm ) ! (z ) in Vl , where q(z ) is a solution of the problem (2.13.12) with q = z . It is easily seen that q0 = z is a

solution of the problem (2.13.29), concluding the proof. In connection with nding a solution of the problem (2.13.29), there arises the question on di erentiability of the functionals i . By using Theorem 2.13.4 and the chain rule, we obtain the following

Theorem 2.13.6 Let the conditions of Theorem 2.13.4 hold, and let i : (q u) ! i (q u) is a Frechet continuously di erentiable mapping of M Vl into R. Then, the functional i de ned by (2.13.27) is a Frechet continuously di erentiable mapping of M into R, and the Frechet derivative 0i of i at a point q 2 M is given
by
i (q )h =
0

@ i (q (q))h + @ i (q (q)) @q @u

(q ) h

h 2 X:

Approximate solution of the optimization problem (2.13.29) is connected with obtaining approximations of the direct problem (2.13.1), (2.13.2) (or (2.13.12)) which converge to the exact solution in the norm of the space Vl . In many applications there also arises the problem of construction of approximate solutions converging to the exact one in a strong norm. Galerkin and Riesz methods provide convergence in weak norms corresponding to generalized solutions, and so these cannot be used in that case. We consider now a method of minimization of residual for the problem (2.13.1), (2.13.2). ^ Let us examine rst the case when Vl contains only zero element and (f g) 2 Hl . Then, by Theorem 2.13.1, there exists a unique solution u of the problem (2.13.1), (2.13.2) and the following estimate holds:

2.13.4 Approximate solution of the direct problem ensuring convergence in the norm of a space of smooth functions

kukVl ck(f g)kHl

(2.13.31)

where k kVl and k kHl are the norms of the spaces Vl and Hl and c is a constant independent of f and g. Let fwi g be a sequence such that 1. wi 2 Vl for all i 2. the elements w1 : : : w are linearly independent for an arbitrary 3. nite linear combinations
P

ci wi (ci 2 R) are dense in Vl .

176

Chapter 2. Optimal control by coe cients in elliptic systems

Let Vl be the linear span of the elements w1 : : : w , i.e.,

Vl = v j v =

De ne an approximate solution u of the problem (2.13.1), (2.13.2) as follows: u 2 Vl kLu ; (f g)kHl = vmin kLv ; (f g)kHl : (2.13.32) 2V One can see that, under the above mentioned assumptions, for each there exists a solution of the problem (2.13.32) and lim ku ; ukVl = 0: (2.13.33) !1 However, (2.13.32) is a di cult problem of minimization of a maximum function. So, from the viewpoint of computation, it is convenient to replace the space Hl with a Hilbert space Hl given by
l

i=1

ci wi ci 2 R :

Hl =
C l;si ( ) H
i(

m Y

q = 1 : : : r (see (2.13.5)), i and q being integers. We suppose that (f g) 2 Hl and the functions wi satisfy Conditions 1, 2, 3 with Vl replaced by Vl1 such that Hl Vl1 . The space Vl1 may be de ned by the imbedding theorem. Then, we can replace the problem (2.13.32) with the following one: Find u satisfying ~ u 2 Vl1 ~ (2.13.34) 2 = min kLv ; (f g )k2 : kLu ; (f g)kHl v2V ~ Hl
This is the problem of minimization of a quadratic functional and it is signi cantly easier than the problem (2.13.32). For an arbitrary there exists a solution of the problem (2.13.34) and it may be shown that the above mentioned conditions imply u ! u in Vl . ~ ^ Suppose now the dimension of the space Vl is not equal to zero and (f g) 2 Hl . By Theorem 2.13.1, there exists a unique solution u of the problem (2.13.1), ^ (2.13.2) such that u 2 Vl . Let f'1 : : : 'k g be a basis in Vl . For > k, de ne el1 subspaces V Vl1 as follows: e Vl1 = v j v 2 Vl1 (v 'i )L2 ( )m = 0 i = 1 : : : k : (2.13.35) De ne an approximate solution u of the problem (2.13.1), (2.13.2) by
l1

i=1

H i( ) C l; q (S )

r Y

q=1

H q (S ) i = 1 ::: m

H q (S )

kLu ; (f g)k2 l = min kLu ; (f g)k2 l : H H e

v2Vl1

e u 2 Vl1

(2.13.36)

2.13. Optimization problems with smooth solutions of state equations

177

The problem (2.13.36) may be solved by, e.g., Lagrange multipliers, and it may be shown that u ! u in Vl .

178

Chapter 2. Optimal control by coe cients in elliptic systems

Chapter 3

Control by the right hand sides in elliptic problems

\ `Let us sit on this log at the road side,' says I, `and forget the inhumanity and ribaldry of the poets. It is in the glorious columns of ascertained facts and legalised measures that beauty is to be found.' " | O. Henry \The Handbook of Hymen"

3.1.1 Setting of the problem. Auxiliary statements

Let U be a re exive Banach space, and suppose

3.1 On the minimum of nonlinear functionals

u v ! (u v) is a bilinear, symmetric, continuous, positive form on U U ,
i.e., (3:1:1)

u v2U j (u v)j ckukU kvkU u v 2 U c = const > 0 ( u u) 0 u2U

(u v) = (v u) Suppose also that

v ! Q(v) is a linear continuous form on U ,

179

(3.1.2)

180

Chapter 3. Control by the right hand sides in elliptic problems

Uad is a convex, closed, bounded set in U .

Consider the quadratic functional

(3.1.3) (3:1:4)

J (v) = (v v) ; 2Q(v)
which is to be minimized on the set Uad. Before considering this problem, we state two auxiliary statements.

Lemma 3.1.1 Let u v ! (u v) be a bilinear, symmetric, continuous, positive form on U U , where U is a Banach space. Then, the function v ! (v v) is
lower semicontinuous in the weak topology of the space U .

Proof. It is easy to verify the following equality (v v) = (u u) + 2 (u v ; u) + (v ; u v ; u) u v 2 U: (3:1:5) Let vn ! v0 weakly in U . Setting in (3.1.5) v = vn , u = v0 , we get (vn vn ) = (v0 v0 ) + 2 (v0 vn ; v0 ) + (vn ; v0 vn ; v0 ): (3:1:6) Since the form (u v) is continuous on U U , the function v ! (v0 v) is
continuous in the weak topology of the space U . Hence,
n!1

lim (v0 vn ; v0 ) = 0:

Making use of this equality, (3.1.6), and of the fact that (u u) 0 for all u 2 U , we obtain lim inf (vn vn ) (v0 v0 ): n!1

Lemma 3.1.2 Let u v ! (u v) be a bilinear, symmetric, continuous form on U U , where U is a Banach space, and let
(u u) 0

u 2 U:

(3:1:7)

Then, the function u ! (u u) is convex, i.e., for all u1 u2 2 U and for all t 2 (0 1), the following estimate is true :

((1 ; t)u1 + tu2 (1 ; t)u1 + tu2 ) (1 ; t) (u1 u1 ) + t (u2 u2 ):

(3:1:8)

If the equality in (3.1.7) occurs only for u = 0, then the function u ! (u u) is strictly convex, i.e., the equality in (3.1.8) takes place only when u1 = u2.

Proof. Introduce the notation

u0 = (1 ; t)u1 + tu2 :
(3:1:9)

3.1. On the minimum of nonlinear functionals

181 (3:1:10) (3:1:11)

By setting in (3.1.5) v = u1 , u = u0 , and taking into consideration (3.1.7), we have (u1 u1 ) In the same way, we get (u2 u2 ) (u0 u0 ) + 2 (u0 u2 ; u0): If the function u ! (u u) vanishes only at u = 0, then the equality in (3.1.10) holds only for u1 = u0, and in (3.1.11) it holds only for u2 = u0 , i.e., if and only if u1 = u2 . By multiplying the inequalities (3.1.10) and (3.1.11) by 1 ; t and t, respectively, where t 2 (0 1), and summing them up, we obtain in view of (3.1.9) that (1 ; t) (u1 u1 ) + t (u2 u2 ) (u0 u0 ) + 2 (u0 (1 ; t)u1 + tu2 ; u0 ) = (u0 u0 ) + 2 (u0 0) = (u0 u0 ): (3.1.12) (u0 u0 ) + 2 (u0 u1 ; u0):

If the condition (u u) = 0 yields u = 0, then in (3.1.12) the equality holds only when u1 = u2. Now, (3.1.8) is implied by (3.1.9) and (3.1.12), concluding the proof.

Theorem 3.1.1 Let the conditions (3.1.1){(3.1.3) hold, and let a functional J (v) be determined by the relation (3.1.4). Then, the subset X de ned by the formula
X = u j u 2 Uad J (u) = v2U J (v) inf
ad

3.1.2 The existence theorem

is nonempty, closed in Uad , and convex. If the function v ! (v v) vanishes only at v = 0, then the set X contains only one element.

Proof. Let fvng Uad be a minimizing sequence, i.e., J (vn ) ! v2U J (v): inf ad
By (3.1.3), we conclude

(3:1:13)

8n: Therefore, we can choose from the sequence fvng a subsequence fvm g such that vm ! w weakly in U: (3:1:14)
Since Uad is a closed, convex set in the Banach space U , by the Mazur theorem (see, e.g., Yosida (1971)), Uad is sequentially weakly closed. Hence, (3.1.14) yields

kvn kU const

w 2 Uad :

(3:1:15)

182

Chapter 3. Control by the right hand sides in elliptic problems

By virtue of (3.1.2), the function v ! Q(v) is continuous in the weak topology of the space U . This and Lemma 3.1.1 imply that the function v ! J (v) is lower semicontinuous in the weak topology of U . Now, due to (3.1.14), we have lim inf J (vm ) J (w): m!1 Then, by (3.1.13) and (3.1.15), we conclude

w 2 Uad
Thus,

J (w) v2U J (v): inf ad

ad

and the set X is nonempty. Let fun g be a sequence of elements from X , and let un ! u in U . Then u 2 Uad J (u) = nlim J (un ) = v2U J (v): inf !1
ad

J (w) = v2U J (v) inf

Hence, u 2 X and the set X is closed in Uad . To prove that X is convex, take u1 u2 2 X then

J (u1 ) = J (u2 ) = v2U J (v): inf

ad

(3:1:16)

By virtue of Lemma 3.1.2, the function v ! (v v) is convex. Hence, the function v ! J (v) is also convex, and if t 2 (0 1), then

J ((1 ; t)u1 + tu2) (1 ; t)J (u1 ) + tJ (u2 ) = v2U J (v): inf

ad

From (3.1.3) consequently,

(1 ; t)u1 + tu2 2 Uad

(1 ; t)u1 + tu2 2 X: Now, let the function v ! (v v) vanish only at v = 0. Let us show that, in this case, the set X contains only one element. Suppose that there exist two elements u1, u2 of the set Uad such that u1 6= u2 and (3.1.16) holds. Since the function v ! (v v) vanishes only at v = 0, by virtue of Lemma 3.1.2 this function is strictly convex. Therefore, the function v ! J (v) is also strictly convex, and 1 1 J 2 (u1 + u2 ) < 2 (J (u1 ) + J (u2 )) = v2U J (v): inf (3:1:17) ad 1 Since the set Uad is convex, we have 2 (u1 + u2) 2 Uad . So, the relation (3.1.17) contains a contradiction. Hence, u1 = u2 , and the theorem is proved.

3.1. On the minimum of nonlinear functionals

183

Remark 3.1.1 Let the above assumptions hold and let the function v ! (v v) meet the condition (v v) ckvk2 v 2 U c = const > 0: (3:1:18) U Then, due to the following inequality J (v) = (v v) ; 2Q(v) ckvk2 ; c1 kvkU U a sequence fvn g satisfying the condition (3.1.13) is bounded in U . Therefore, if Uad is a closed, convex set in U (which is not necessarily bounded), then there exists a function u 2 Uad such that J (u) = v2U J (v): inf (3:1:19)
ad

Moreover, due to (3.1.18), the function v ! (v v) vanishes only at v = 0. So, there exists a unique element u 2 Uad satisfying (3.1.19). Remark 3.1.2 Let U be endowed with the strong topology, and let Uad be a compactum in U (which is not necessarily convex). By (3.1.1), (3.1.2), and (3.1.4), the function v ! J (v) is continuous in the strong topology of the space U , so that there exists an element u 2 Uad that satis es the condition (3.1.19). Remark 3.1.3 In the proof of Theorem 3.1.1, we needed the closedness and convexity of the set Uad only to derive (3.1.15) from (3.1.14). Therefore, the statement of Theorem 3.1.1 remains valid if we assume that the set Uad is bounded in the strong topology of U and sequentially weakly closed, instead of the condition (3.1.3).The sequential weak closedness means that, if fvn g Uad , vn ! v weakly in U , then v 2 Uad (see subsec. 1.2.3).

Theorem 3.1.2 If the assumptions of Theorem 3.1.1 hold, then an element u of

3.1.3 Characterization of a minimizing element

Uad belongs to the set X if and only if the following inequality holds (u v ; u) Q(v ; u) v 2 Uad: (3:1:20) Proof. Let u 2 X . Since the set Uad is convex, for all v 2 Uad and t 2 (0 1) we
have whence

J (u) J ((1 ; t)u + tv)

J (u + t(v ; u)) ; J (u) 0 t 2 (0 1): (3:1:21) t The functional J (v) being de ned by the equality (3.1.4), we easily conclude that, for all w h 2 U , the following relation is valid lim J (w + tht) ; J (w) = J 0 (w)h = 2 (w h) ; Q(h)]: t!0

184

Chapter 3. Control by the right hand sides in elliptic problems

Here, J 0 (w) is the linear continuous form on U given by

y ! J 0 (w)y = 2 (w y) ; Q(y)]:

(3:1:22)

So, in (3.1.21), we can pass to the limit as t tends to zero from the right. Thus, we get J 0 (u)(v ; u) 0 v 2 Uad: (3:1:23) Therefore, from (3.1.22) we deduce (3.1.20). Conversely, let the relation (3.1.20) hold (so that (3.1.23) holds, too). By Lemma 3.1.2, the function v ! J (v) is convex, and so for any t 2 (0 1)

obtain

J ((1 ; t)w + tv) ; J (w) v w 2 U: t Passing to the limit as t ! 0 on the right hand side of this inequality, we J (v) ; J (w) J (v) ; J (w) J 0 (w)(v ; w): v 2 Uad

Setting here w = u and taking into account (3.1.23), we deduce that

J (v) ; J (u) 0
concluding the proof.

Remark 3.1.4 Suppose that (u v) is a bilinear, symmetric, continuous, coercive form on U U , where U is a Hilbert space over R. Further, let v ! Q(v) be a linear continuous form on U , and let Uad be a closed, convex cone with peak at the origin of coordinates. Then, there exists a unique function u 2 Uad satisfying the
conditions (3.1.19) and (3.1.20) (see Remark 3.1.1). Moreover, (3.1.20) is equivalent to the following relations (u v) Q(v)

v 2 Uad

(u u) = Q(u):

Indeed, since Uad is a cone, u + v 2 Uad for all v 2 Uad. By substituting in (3.1.20) u + v for v, we get the rst inequality. By setting v = 0 in (3.1.20), we obtain (u u) Q(u), hence, (u u) = Q(u). Conversely, adding the equality (u ;u) = Q(;u) to the inequality (u v) Q(v), we have (3.1.20).

3.1.4 Functionals continuous in the weak topology

Uad U , Uad is bounded in the strong topology of U and sequentially weakly closed.

Hitherto, v ! J (v) was assumed to be a quadratic functional. Below, we will present a statement in which the functional J (v) is not supposed to be quadratic. Assume, as before, that U is a re exive Banach space, and (3:1:24)

3.2. Approximate solution of the minimization problem

185

Suppose also that

v ! J (v) is a lower semicontinuous mapping of Uad into R =

with respect to the topology generated by the weak one of the space U .

(3:1:25)

Theorem 3.1.3 If the conditions (3.1.24) and (3.1.25) are ful lled, then there exists a function u such that
u 2 Uad J (u) = v2U J (v): inf
ad

(3:1:26) (3:1:27)

Proof. Let fvng be a minimizing sequence, i.e., vn 2 Uad 8n lim J (v ) = inf J (v): n!1 n v2Uad
By (3.1.24), we conclude

kvn k const 8n: Hence, we can choose from the sequence fvn g a subsequence fvm g such that vm ! w weakly in U: (3:1:28)
Since Uad is sequentially weakly closed by the hypothesis, (3.1.28) yields

w 2 Uad :
By combining (3.1.25), (3.1.27), and (3.1.28), we conclude
v2Uad

(3:1:29)

inf J (v) = lim inf J (vm ) J (w): m!1

Therefore, from (3.1.29) we deduce the function u = w to satisfy the relation (3.1.26).

3.2 Approximate solution of the minimization problem

3.2.1 Inner point lemma
Below, we will need the following

Lemma 3.2.1 Let A be a convex set in the Banach space U which contains at least one inner point y0 , and let A be the closure of A. If y 2 A, then every point
of the open interval with the ends y0 and y is an inner point of the set A.

186 i.e.,

Chapter 3. Control by the right hand sides in elliptic problems

Proof. Let z be an arbitrary point of the open interval with the ends y and y0,
z = (1 ; t)y + ty0 = y + t(y0 ; y) t 2 (0 1): Further, let g ! f (g) be a mapping of U into U de ned by the relation f (g) = z + (g ; z )
where (3:2:1) (3:2:2) (3:2:3)

Obviously, g ! f (g) is continuous the inverse mapping is given by y ! f ;1 (y) = 1 (y + ( ; 1)z ) and maps U into U continuously, too. Hence,

t = ;1 ; t:

g ! f (g) is a homeomorphism of U onto U .

Taking into consideration (3.2.1), one can easily see that

(3:2:4) (3:2:5)

f (y0) = y:

If Q is an open set in U containing the point y0 and belonging to A, then, by virtue of (3.2.4), f (Q) is an open set in U , y 2 f (Q) (because of (3.2.5)), and there exists a point h such that

h2Q

f (h) 2 A: q 2 A:

Indeed, if > 0 and su ciently small, then the following inclusion holds

q = y + (y0 ; y) 2 f (Q)
So, we can take h = f ;1 (q). By (3.2.2), we get

f (h) ; z = (h ; z ) = (h ; f (h)) + (f (h) ; z ): z = f (h) + ; 1 (h ; f (h)) i.e., z is the image of the function h under the mapping g ! '(g) = f (h) + ; 1 (g ; f (h)):
Hence, (3:2:6) (3:2:7)

3.2. Approximate solution of the minimization problem

187

Since < 0 in view of (3.2.3), we get 0 < ; 1 < 1: (3:2:8) From this we conclude that the function g ! '(g) is a homeomorphism of U onto U , and so '(Q) is an open set in U: (3:2:9) Since h 2 Q, from (3.2.6) we deduce that

z 2 '(Q): (3:2:10) Let g be an arbitrary point from Q. Since f (h) 2 A and the set A is convex, we conclude that '(g) 2 A from (3.2.7) and (3.2.8). Thus, '(Q) A: (3:2:11) Finally, by (3.2.9){(3.2.11), we conclude z to be an inner point of the set A.
The lemma is proved.

3.2.2 Finite-dimensional problem

n

Below we suppose U to be a re exive, separable Banach space. Then, there exists a sequence of nite-dimensional subspaces fUng U which satis es the limit density condition: lim inf kv ; wkU = 0 w 2 U: (3:2:12) n!1 v2U The problem of minimization of the functional J (v) on Uad is replaced by the following one: Find a function un such that

un 2 Un \ Uad

J (un ) = v2Uinf U J (v): \

n
ad

(3:2:13)

Finite-dimensional problem for the quadratic functional Theorem 3.2.1 Let U be a re exive, separable Banach space, let the assumptions of Theorem 3.1.1 be satis ed, and let fUng be a sequence of nite-dimensional

subspaces of U meeting the relation (3.2.12). Suppose that the set Uad is endowed with the topology generated by the strong one of the space U , and in this topology the set Uad has at least one inner point. Then, for all n su ciently large, the problem (3.2.13) has a solution un and
n!1

lim J (un ) = v2U J (v): inf

ad

(3:2:14)

One can choose a subsequence fumg of the sequence fung such that um ! u weakly in U , and u 2 X . If the set X consists of one element u only, then the sequence fung weakly converges to u in U .

188

Chapter 3. Control by the right hand sides in elliptic problems

Proof. Let u 2 X , i.e.,

u 2 Uad

J (u) = v2U J (v): inf

ad

(3:2:15)

Denote by U ad the interior of Uad. By the hypothesis, there exists a function g belonging to U ad . Let fti g1 be a sequence such that ti 2 (0 1) for all i, and limi!1 ti = 0. i=1 From Lemma 3.2.1, we deduce that

zi = u + ti (g ; u) 2 U ad
and Let " be an arbitrary positive number. In view of (3.2.16), we can choose i = i" such that " kzi" ; ukU 2 : (3:2:17) 0< where 2 lim kz ; ukU = 0: i!1 i (3:2:16)

Since zi" 2 U ad , there exists a constant such that

"

d ( zi" ) Uad :

(3:2:18)

satisfying

d ( zi" ) = v v 2 U kv ; zi" kU By (3.2.12), there is l > 0 such that, for each n

l, there exists gn 2 Un

(3:2:19) (3:2:20)

From (3.2.17), (3.2.18), and (3.2.20), we deduce that kgn ; ukU kgn ; zi" kU + kzi" ; ukU " n l and the relations (3.2.18){(3.2.20) yield that gn 2 Uad . From the above argument, we conclude the existence of a sequence fgn g1 l n= such that gn 2 Un \ Uad lim kg ; ukU = 0: (3:2:21) n!1 n Hence, lim J (g ) = J (u) = v2U J (v): inf (3:2:22) n!1 n When n l, the set Un \ Uad is nonempty. By (3.1.3), Un \ Uad is a compactum, and since v ! J (v) is a continuous mapping of U into R, the problem (3.2.13) has a solution un. Further, (3.2.13), (3.2.15), and (3.2.21) imply that
ad

kgn ; zi" kU

J (gn ) J (un ) J (u):

3.2. Approximate solution of the minimization problem

189

Combining this with (3.2.22), we get (3.2.14). By virtue of (3.1.3), the sequence fun g is bounded in U . Let us choose a subsequence fumg from it such that um ! w weakly in U: Then, w 2 Uad , because Uad is a convex set that is closed in the strong topology of U . Taking into account that the mapping v ! J (v) is lower semicontinuous in the weak topology of U , and using (3.2.14), we have lim inf J (um ) = mlim J (um ) = v2U J (v) J (w): inf m!1 !1

J (w) = v2U J (v) inf ad i.e., w 2 X , and we can set u = w. Suppose that the set X contains only one element u. Let us show that then un ! u weakly in U: (3:2:23) Assume that (3.2.23) is not true. Then, there exist an element q 2 U , a constant " > 0, and a subsequence fuk g of the sequence fung such that j(q uk ; u)j " 8k: (3:2:24) Since fuk g Uad , the subsequence fuk g is bounded. Let us choose a subsequence fui g from it such that ui ! y weakly in U: (3:2:25)
Therefore, from (3.2.14) we get lim inf J (ui ) = ilim J (ui ) = v2U J (v) J (y): inf i!1 !1
ad

Thus,

ad

The set Uad being sequentially weakly closed, from (3.2.25) we deduce y 2 Uad . Consequently, y 2 X and y = u because the set X contains only one element u. Thus, the formulas (3.2.24) and (3.2.25) contain a contradiction. The theorem

Remark 3.2.1 Let, under the assumptions of Theorem 3.2.1, U be a Hilbert space, and let a form (v w) be coercive, i.e., (v v) ckvk2 v 2 U c = const > 0: U Then the form (v w) generates in U a scalar product and a norm equivalent to the original one of the space U . Moreover, there exists a unique element u meeting (3.2.15) (see Remark 3.1.1), and (3.2.14) implies that (un un) ! (u u): Therefore, un ! u weakly in U yields that un ! u strongly in U .

is proved.

190

Chapter 3. Control by the right hand sides in elliptic problems

Example

Let be a bounded open set in Rm , x = fx1 : : : xm g 2 , dx = dx1 : : : dxm , U = L2 ( ). For v w 2 L2 ( ), we set (v w) =

Z

vw dx

v ! Q(v) =

gv dx
)

g 2 L2 ( ): c

Obviously, the form is coercive. Let the set Uad be of the form

Uad = v j v 2 L2( ) kvkL2(

c being a positive number. On the set Uad, we introduce the topology generated by that of the space L2 ( ). Then, Uad is a bounded, closed, convex set in U = L2 ( ) with a nonempty interior U ad . So, Theorems 3.1.1 and 3.2.1 can be applied. Suppose now that the set Uad is determined by the following relation Uad = v j v 2 L2 ( ) c0 v(x) c1 a.e. in c0 < c1 c0 and c1 are constants : (3.2.26) U = L2 ( ).
It is easy to see that Uad is a closed, bounded, convex subset of the space In this case, we can use Theorem 3.1.1. From this theorem and Remark 3.1.1 we deduce the existence of a unique function u such that

u 2 Uad
where

J (u) = v2U J (v) inf

ad

(3:2:27) (3:2:28)

J (v ) =

v2 dx ; 2

gv dx:

However, Theorem 3.2.1 cannot be applied, because the set Uad de ned by (3.2.26) is \thin," that is, the interior U ad of Uad in the topology generated by the strong one of the space L2 ( ) is empty. We proceed as follows. Let fHn g be a sequence of nite-dimensional subspaces of L2( ) such that the set Hn \ Uad is nonempty for all n. Then, by the coercivity of the form , for each n there exists a unique function un such that

un 2 Hn \ Uad gn 2 Hn \ Uad

J (un ) = H inf J (v): \U

n
ad

If we prove the existence of a sequence fgn g such that

gn ! u strongly in L2 ( )

(3:2:29)

3.2. Approximate solution of the minimization problem

191

where u satis es the conditions (3.2.27), then by Theorem 2.4.3 we will conclude that lim J (un ) = J (u) = v2U J (v): inf (3:2:30) n!1 Moreover, the boundedness of the set Uad will imply that un ! u weakly in L2 ( ), and since (3.2.30) yields that
ad

kunkL2( ) ! kukL2(

we get that un ! u strongly in L2( ). Let us construct the subspaces Hn for the problem under consideration. To every index n we set a corresponding partition of in a nite number of disjoint subsets 1n 2n : : : Nnn which are Lebesgue measurable. Denote such a partition by n . For every set in , i = 1 2 : : : Nn , de ne its diameter
in = sup kx ; y kRn x y2 in

and let n = 1 max in . We suppose that i Nn lim = 0: n!1 n Denote by in the characteristic function of the set in :
in (x) =
(

(3:2:31)

1 if x 2 in 0 if x 62 in :

De ne now subspaces Hn by

Hn = v j v =

Nn X i=1

ci in ci are constants :

(3:2:32)

The subspaces Hn satisfy the limit density condition in L2 ( ) (see, e.g., Schwartz (1967)), moreover, for every function w 2 Uad , where Uad is de ned by (3.2.26), there exists a sequence fwn g such that

wn 2 Hn \ Uad

wn ! w strongly in L2 ( ):

Thus, if one solves the problem (3.2.13) with the functional (3.2.28), with the set Uad determined by the formula (3.2.26), and with Hn de ned by (3.2.32), (3.2.31), then (3.2.30) holds and un ! u strongly in L2 ( ).

192

Chapter 3. Control by the right hand sides in elliptic problems

Now, we are going to examine the case when the set Uad and the functional J (v) meet the conditions (3.1.24) and (3.1.25). By virtue of Theorem 3.1.3, there exists a function u such that u 2 Uad J (u) = v2U J (v): inf (3:2:33)
ad

Nonquadratic functional

As before, we will denote by U ad the interior of the set Uad , provided Uad is equipped with the topology generated by the U -strong one. Theorem 3.2.2 Let U be a re exive, separable Banach space, let fUng be a sequence of nite-dimensional subspaces of U which satisfy the relation (3.2.12), and let Uad meet the condition (3.1.24). Assume that the functional J (v) satis es the condition (3.1.25), and, moreover, v ! J (v) is a continuous mapping of Uad into R with re(3:2:34) spect to the topology generated by the U -strong one. Let also there exist a sequence fqn g such that

qn 2 U ad 8n

qn ! u strongly in U

(3:2:35)

From the sequence fung one can choose a subsequence fum g such that um ! u weakly in U . If there exists a unique function u for which (3.2.33) is valid, then un ! u weakly in U .

where u is the function satisfying (3.2.33). Then, for su ciently large n, the problem (3.2.13) has a solution un , and lim J (un ) = J (u) = v2U J (v): inf (3:2:36) n!1
ad

Therefore, from (3.2.34) we have lim J (g ) = J (u) = v2U J (v): inf n!1 n
ad

Proof. Analogously to the proof of Theorem 3.2.1, by using (3.2.12) and (3.2.35), one can show that there exists a sequence fgn g1 k , k being a positive number, n= such that gn 2 Un \ Uad gn ! u strongly in U: (3:2:37)
(3:2:38)

When n k, the set Un \ Uad is not empty, and in view of (3.1.24) it is a compactum. So, (3.2.34) implies that the problem (3.2.13) has a solution un for any n k. By (3.2.13), (3.2.33), and (3.2.37), we deduce that J (gn ) J (un ) J (u):

3.3. Control by the right hand side in elliptic problems provided

:::

193

This inequality together with (3.2.38) yields (3.2.36). By virtue of (3.1.24), the sequence fun g is bounded. Let us choose a subsequence fum g from it such that

um ! w

weakly in U:

(3:2:39) (3:2:40)

Hence, taking into account (3.1.25) and (3.2.36), we have lim inf J (um ) = mlim J (um ) = v2U J (v) J (w): inf m!1 !1
ad

From (3.1.24) and (3.2.39) it follows that that w 2 Uad . Now, by (3.2.40), we conclude J (w) = v2U J (v) inf
ad

and the function u = w meets the relations (3.2.33). If there exists a unique function u that satis es (3.2.33), then one can conclude the convergence un ! u weakly in U in the same way as in the proof of Theorem 3.2.1.

Remark 3.2.2 In Theorems 3.2.1 and 3.2.2, we supposed that U ad was not empty.

If U ad is an empty set, then to investigate the solvability of the problem (3.2.13) and the convergence of its solutions, one can apply Theorem 2.4.3. In this case, the proof of the existence of a sequence fgng which meets (2.4.28) is the most serious obstacle when this theorem is used. However, as have been seen in the example of the present subsection, when the set Uad has a comparatively simple structure, one does manage to apply Theorem 2.4.3.

3.3 Control by the right hand side in elliptic problems provided the goal functional is quadratic
3.3.1 Setting of the problem
Let H be a Hilbert space over R and let

H that is symmetric, (3:3:1) continuous, and coercive. Then, for a given element f 2 H , there exists a unique u 2 H such that a(u v) = (f v) v2H
(3:3:2) and f ! u is a linear continuous mapping of H into H . Suppose we are given a set of controls U such that

a(w v) be a bilinear form on H

U is a re exive Banach space

194 and

Chapter 3. Control by the right hand sides in elliptic problems

(3:3:3) Suppose that some system is described by the form a(w v), i.e., for every control g 2 U , the state of the system u 2 H is determined as the solution of the problem a(u v) = (f + Bg v) v 2 H: (3:3:4) It is obvious that u depends on g (the element f is assumed to be xed), so that we will write u(g). Then a(u(g) v) = (f + Bg v) v 2 H: (3:3:5) In view of (3.3.1), the equation (3.3.5) determines uniquely the state u(g) Moreover, suppose we are also given an observation z (g) = Lu(g) where L 2 L(H H), H is a Hilbert space. To every control g 2 U there corresponds the value of the goal functional J (g) = kLu(g) ; z0 k2 (3:3:6) H z0 being a given element of the space H. About the set of admissible controls we suppose that Uad is a convex, closed, bounded set in the space U . (3:3:7) It should be stressed that here the closedness and boundedness is considered in the strong topology of U as we agreed in subsec. 1.2.3, if the precise reference to a topology of a normed space is omitted, then we mean the strong one. The optimal control problem consists in nding y such that y 2 Uad J (y) = g2U J (g): inf
ad

B 2 L( U H ) :

Due to (3.3.5), g ! u(g) is an a ne mapping of U into H , and a((u(g) ; u(0)) + u(0) v) = (f + Bg v) v 2 H: Therefore, a(u(g) ; u(0) v) = (Bg v) v 2 H: Hence, g ! (u(g) ; u(0)) is a linear continuous mapping of U intoo H.

3.3.2 Existence of a solution. Optimality conditions

(3:3:8)

3.3. Control by the right hand side in elliptic problems provided

:::

195

Rewrite J (g) as

J (g) = kL(u(g) ; u(0)) + Lu(0) ; z0 k2 : H

If, for arbitrary elements y g from U , we set (y g) = (L(u(y) ; u(0)) L(u(g) ; u(0)))H (F g) = (z0 ; Lu(0) L(u(g) ; u(0)))H then (3.3.9) (3.3.10)

J (g) = (g g) ; 2(F g) + kz0 ; Lu(0)k2 : (3:3:11) H Since L 2 L(H H), (3.3.8){(3.3.10) imply that F 2 U and that y g ! (y g) is a bilinear, symmetric, continuous, positive form on U U .
By applying Theorem 3.1.1, we obtain the following

Theorem 3.3.1 Let the conditions (3.3.1){(3.3.3) be ful lled, and let the state of the system be determined as the solution of the problem (3.3.4), where f is a xed element of H , g 2 U . Let also a goal functional J be de ned by the relation (3.3.6), where z0 2 H, L 2 L(H H), and let the set Uad meet the condition (3.3.7).
Then, the subset X de ned by the relation

X = y j y 2 Uad J (y) = g2U J (g) inf

ad

is nonempty, closed in Uad , and convex. If the function g ! (g g) determined by the formula (3.3.9) vanishes only at g = 0, then the subset X contains only one element.

Remark 3.3.1 Assume that the above suppositions hold, U is a nite-dimensional subspace of H , B is an embedding of U into H , and the function g ! (g g) vanishes only at g = 0. Then, the form (y g) determines a scalar product and a norm in U . Since in a nite-dimensional vector space every two norms are equivalent, the bilinear form (y g) is coercive in U . Hence, Remark 3.1.1 yields that if Uad is a closed, convex set in U (not necessarily bounded), then the subset X contains only one element.
Now, let us apply Theorem 3.1.2 to the functional (3.3.6) or, equivalently, to (3.3.11). If y 2 X , then this theorem together with (3.3.9){(3.3.11) yields 1 J 0 (y)(g ; y) = (y g ; y) ; (F g ; y) 2 = (L(u(y) ; u(0)) L(u(g ; y) ; u(0)))H ; (z0 ; Lu(0) L(u(g ; y) ; u(0)))H = (Lu(y) ; z0 L(u(g ; y) ; u(0)))H 0 g 2 Uad: (3.3.12)

196

Chapter 3. Control by the right hand sides in elliptic problems

By (3.3.5), we conclude a(u(g) v) = (f + Bg v) v2H a(u(y) v) = (f + By v) v 2 H: Therefore, a(u(g) ; u(y) v) = (B (g ; y) v) v 2 H: (3:3:13) Further, we have a(u(g ; y) v) = (f + B (g ; y) v) v 2 H: From this equality, taking into account that a(u(0) v) = (f v) v2H we obtain a(u(g ; y) ; u(0) v) = (B (g ; y) v) v 2 H: (3:3:14) In view of (3.3.1), the bilinear form a(w v) generates a scalar product and a norm in the space H , which is equivalent to the original norm of the space H . So, by (3.3.13) and (3.3.14), we get u(g ; y) ; u(0) = u(g) ; u(y): Hence, we can represent the optimality condition (3.3.12) in the form (Lu(y) ; z0 L(u(g) ; u(y)))H 0 g 2 Uad: (3:3:15) Thus, we have proved the following Theorem 3.3.2 Let the conditions of Theorem 3.3.1 hold. Then, the function y belongs to the set X if and only if the inequality (3.3.15) holds. Notice that in order to get an approximate solution of the problem of minimization of the functional (3.3.6) on the set Uad that satis es the condition (3.3.7), one can apply Theorem 3.2.1 in the case when the set Uad has at least one interior point. If it is not the case, one can use Theorem 2.4.3 and the reasoning from the example in subsec. 3.2.2. Nevertheless, it is better to choose a set of controls U in such a way that the set Uad contain interior points (see also Remark 3.2.2 and the example below).

3.3.3 An example of a system described by the Dirichlet problem

Direct problem
1 Let H = H0 ( ), where

be a bounded domain in Rn , and let

n XZ i j =1

a(w v) =

@w @v dx + Z a wv dx aij @x @x 0 i j

(3:3:16)

3.3. Control by the right hand side in elliptic problems provided

:::

197

where

aij (x) i j i j =1 const > 0, a0 (x)

From (3.3.16){(3.3.18)

n X

aij a0 2 L1 ( )

n X 2 i a.e. in i=1

aij = aji
for all

2 Rn ,

9 > == >

(3.3.17) (3.3.18)

> 0 a.e.in .

a(v v)
Thus, from (3.3.17) we get

kvk2 1 ( ) : H
(3:3:19)

the bilinear form a(w v) determined by (3.3.16) is contin1 1 uous, symmetric, and coercive on H0 ( ) H0 ( ).
;

1 Hence, if q 2 H0 ( ) = H ;1( ), then there exists a unique element u 2 1 ( ) such that H0 1 a(u v) = (q v) v 2 H0 ( ): (3:3:20) De ne the elliptic operator A of the second order:

Aw = ;

@ a @ w + a w: ij @x 0 i i j =1 @xj
1 u 2 H0 ( ):

n X

Then, the equation (3.3.20) is equivalent to the following one

Au = q a(u ') = (q ')

(3:3:21)

1 Indeed, since D( ) is dense in H0 ( ), (3.3.20) is equivalent to the equation

' 2 D( )

which, in turn, is equivalent to the equation (3.3.21) by the de nition of a distribution derivative. The problem (3.3.21) is called the Dirichlet problem.

Optimization problem
Suppose

U H ;1 ( ), the imbedding U ! H ;1 ( ) is continuous,> > B is the operator of the imbedding of the space U into= 1 H ;1 ( ), H = L2 ( ), L is the imbedding of the space H0 ( )> > into L2 ( ).

(3:3:22)

198

Chapter 3. Control by the right hand sides in elliptic problems

Thus, the state of the system u(g) is determined as the solution of the problem

a(u(g) v) = (f + g v)
is a function de ned on :

1 1 v 2 H0 ( ) u(g) 2 H0 ( )

(3:3:23)

f being a xed function from H ;1( ). The solution u(g) of the problem (3.3.23) x ! u(x g):

Under the conditions (3.3.22), the goal functional is of the form

J (g) = (u(x g) ; z0 (x))2 dx:

By (3.3.9), to this functional there corresponds the bilinear form (y g) =
Z

(3:3:24)

(u(x y) ; u(x 0))(u(x g) ; u(x 0)) dx:

(3:3:25)

R vanishes if and only if g = 0.

Let us show that the function g ! (g g) considered as mapping of U into By (3.3.25), we conclude that (0 0) = 0. Suppose (g g) = 0. Then

u(x g) = u(x 0)

(3:3:26)

and from the relations (3.3.19) and (3.3.26), by using the Riesz theorem, we deduce g = 0. Now, Theorem 3.3.1 implies that, if the set Uad meets the condition (3.3.7), then there exists a unique function y such that

y 2 Uad

(u(y) ; z0

)2 dx =

g2Uad

inf

(u(x g) ; z0 )2 dx:

(3:3:27)

Let us examine some examples of the spaces U and the set Uad for our problem. Example I. Let

U = L2 ( ) Uad = g j g 2 L2 ( ) 0 (x) g 1 (x) a.e. in

(3.3.28)

where 0 and 1 are xed functions from L1 ( ). It is easy to see that the condition (3.3.7) holds and the imbedding U ! H ;1 ( ) is continuous. Hence, there exists a unique function y satisfying (3.3.27). However, since functions from L2 ( ) are not, in general, bounded almost everywhere in , even in the case when
1 (x) ; 0 (x)

a.e. in

c = const > 0

(3:3:29)

3.3. Control by the right hand side in elliptic problems provided

:::

199

the set Uad has no interior points. Example II. Let U = Wpl ( ) p > 1 pl > n l ( ) kg k l Uad = g j g 2 Wp Wp ( ) c1 c1 = const > 0 0 (x) g (x) 1 (x) x 2

(3.3.30) (3.3.31)

where, just as above, 0 , 1 are xed functions from L1( ). The condition (3.3.7) holds and the imbedding U ! H ;1 ( ) is continuous. So, for the space Uad determined by the relations (3.3.30) and (3.3.31), there exists a unique function y satisfying (3.3.27). Since Wpl ( ) C ( ) (by (3.3.30) and the imbedding theorem), under the conditions (3.3.29){(3.3.31), the set Uad has a nonempty interior U ad , so that one can apply Theorem 3.2.1. Notice that, if the condition (3.3.29) is not satis ed, then, by setting 0 1 = 1+" " being arbitrarily small positive number, we have 0 a.e. in : 1 (x) ; 0 (x) " 0 ~ Thus, the set Uad determined by the right hand side of (3.3.31) with 1 substituted for 1 has a nonempty interior. It should be noted that we have enlarged the set Uad , by passing from the ~ \thin" set Uad to its extension Uad , which does have a nonempty interior. Under such an extension, in the case when the topology on Uad is good enough, the set ~ Uad changes slightly (in the example considered, the deviation of Uad from Uad is determined by the parameter ", which can be taken as small as desired). Now let us enlarge the set Uad de ned by the relations (3.3.28) and (3.3.29), c for example, up to a set Uad which has the interior point p = 0 + 2 :

Uad = g j g = 0 + 1 c + q kqkL2( ) " " > 0 : 2 In this case, since functions from L2 ( ) are unbounded in , for every positive " functions g from Uad may take arbitrarily large values, so that they cannot be bounded by functions 0 1 2 L1 ( ).

Another goal functional

;

Let U be a closed subspace of H ;1( ) with the norm of H ;1 ( ), let B be the operator of the imbedding of the space U into H ;1 ( ), H = (L2 ( ))n+1 ,
1 L 2 L H0 ( ) (L2 ( ))n+1

@v @v L : v ! Lv = v @x : : : @x : 1 n

200

Chapter 3. Control by the right hand sides in elliptic problems

De ne a norm in the space H by

v = fv0 v1 : : : vn g

kvk2 = H

n X i=0

bi kvi k2 2 ( L

(3.3.32)

bi = const > 0:
n X Z J (g) = bi i=1

The state of the system u(g) is determined as the solution of the problem (3.3.23), and according to (3.3.6) the goal functional has the form

@u(g) ; z 2 dx + b Z (u(g) ; z )2 dx: 0i 0 00 @xi

(3:3:33)

Here z0i are given elements of L2( ), i = 0 1 : : : n. By (3.3.9), to the goal functional (3.3.33) there corresponds the bilinear form (y g) =
n X Z bi i=1

@u(y) ; @u(0) @xi @xi

+ b0

@u(g) ; @u(0) dx @xi @xi

(u(y) ; u(0))(u(g) ; u(0)) dx: (3.3.34)

The bilinear form (y g) is obviously symmetric and continuous on U U . By (3.3.32) and (3.3.34), we obtain (g g) cku(g) ; u(0)k2 0 ( H1 Since
)

g 2 U c = const > 0:
1 v 2 H0 ( )

(3:3:35)

a(u(g) ; u(0) v) = (g v)
)

from (3.3.19) and the Riesz theorem we conclude

ku(g) ; u(0)kH01 (

c1 kgkH ;1 (

g 2 U c1 = const > 0:

Combining this with (3.3.35), we get the form (y g) to be coercive, i.e., there exists a positive number c2 such that (g g) c2 kgk2 ;1 ( H
)

g 2 U:

Now, by Remark 3.1.1, we get that, if Uad is a closed, convex set in U , which is not necessarily bounded, then there exists a unique function y 2 Uad such that

J (y) = g2U J (g): inf

ad

3.4. Minimax control problems

201

3.4 Minimax control problems

Let H be a Hilbert space over R, let the conditions (3.3.1){(3.3.3) hold, and let the state of the system u(g) be de ned as a solution of the problem (3.3.5), where f is a xed element of H and g 2 U . Suppose we are given mappings Pk : H U ! R such that v q ! Pk (v q) is a lower semicontinuous mapping of H 9 > U into R with respect to the topology generated by the= (3:4:1) product of the weak topologies of H and U , k 2 I , I => f1 2 : : : lg. De ne mappings Qk : U ! R by

g ! Qk (g) = Pk (u(g) g)
and let a goal functional be of the form

k2I

(3:4:2) (3:4:3) (3:4:4)

J (g) = max Qk (g): k2I

Suppose that Uad U , Uad is bounded in the strong topology of U and sequentially weakly closed.

Theorem 3.4.1 Let the conditions (3.3.1){(3.3.3) hold, and let the state of the system u(g) be determined as the solution of the problem (3.3.5), where f is a xed element of H and g 2 U . Let also a goal functional J (g) be de ned by the formulas (3.4.1){(3.4.3), and let the set Uad meet the condition (3.4.4). Then, there exists a function y such that y 2 Uad J (y) = g2U J (g): inf (3:4:5)
ad

To prove this statement, we need the following lemma.

Lemma 3.4.1 Let the conditions (3.3.1){(3.3.3) hold, and let the state of the system u(g) be determined as the solution of the problem (3.3.5), where f is a xed element of H and g 2 U . Then, g ! u(g) is a continuous mapping of the space U endowed with the weak topology into the space H equipped with the weak topology. Proof. Let g 2 U and let fgng be a sequence of elements of U such that gn ! g weakly in U: (3:4:6)
From (3.3.5)

a(u(gn) ; u(g) v) = (B (gn ; g) v)

v 2 H:

(3:4:7)

202

Chapter 3. Control by the right hand sides in elliptic problems

For a xed v 2 H , the function q ! (Bq v) is a linear continuous mapping of U into R. So, (3.4.6) yields (B (gn ; g) v) ! 0: Combining this with (3.4.7), we get
n!1

lim a(u(gn ) ; u(g) v) = 0

v 2 H:

Taking to notice the Riesz theorem, from the latter relation and (3.3.1), we deduce that u(gn ) ! u(g) weakly in H: Thus, the lemma is proved. Proof of Theorem 3.4.1. Let fqng be a minimizing sequence, i.e.,

fqng Uad
(3.4.4) implies

J (qn ) ! g2U J (g): inf

ad

(3:4:8)

kqn kU const 8n: Therefore, from the sequence fqn g we can choose a subsequence fqmg such that qm ! q weakly in U (3:4:9)
moreover, (3.4.4) yields that (3:4:10) By using Lemma 3.4.1, (3.4.1), and (3.4.2), we deduce that g ! Qk (g) is a lower semicontinuous mapping of U equipped with the weak topology into R. Theorem 1.4.3 implies g ! J (g) = max Qk (g) is a lower semicontinuous mapping k 2I (3:4:11) of the space U endowed with the weak topology into R. By (3.4.9) and (3.4.11), we get lim inf J (qm ) J (q): m!1 Hence, from (3.4.8) we get lim inf J (qm ) = g2U J (g) J (q): inf m!1
ad

q 2 Uad :

(3:4:12)

By (3.4.12), taking into account (3.4.10), we obtain

J ( q ) = g 2 U J (g ): inf
ad

3.4. Minimax control problems

203

Hence, the function y = q is a solution of the problem (3.4.5), proving the theorem. Example. Assume that H = H01( ), is a bounded domain in Rn , the form a(w v) is determined by the relations (3.3.16){(3.3.18), U H ;1 ( ), the imbedding U ! H ;1 ( ) is continuous, and B is the imbedding of U into H ;1( ). The state of the system u(g) is de ned as the solution of the problem: Find 1 u(g) 2 H0 ( ) such that

a(u(g) v) = (f + g v) where f is a xed element of H ;1 ( ), g 2 U . g ! Qki (g) =

2 @u(g) ; ki dx @xk Z i g ! Q0i (g) = (u(g) ; 0i )2 dx:
i

1 v 2 H0 ( )

Let be partitioned into a nite number of measurable subsets i , i.e., S and = m i . De ne functions i i=1
Z

k = 1 ::: n i = 1 ::: m

(3.4.13)

Here, ki , 0i are given functions from L2 ( i ). Assign a goal functional through the relation

J (g) = (k max Y Qki (g) i)2I

where I = f0 1 2 : : : ng and Y = f1 2 : : : mg. Lemma 3.4.1 implies that g ! u(g) is a continuous mapping of the space U 1 equipped with the weak topology into H0 ( ) endowed with the weak topology. Making use of Lemma 3.1.1, it is easy to see that
2 @u ; = the function u ! ki dx is lower semicontini @xk 1 uous in the weak topology of H0 ( ).

(3:4:14)

The imbedding theorem yields that

Z

u ! (u ; 0i )2 dx is a continuous mapping of the space= i 1 H0 ( ) endowed with the weak topology into R. Uad meets the condition (3.4.4), then there exists a function y such that y 2 Uad J (y) = g2U J (g): inf
ad

(3:4:15)

Now, Theorem 3.4.1 and the relations (3.4.13){(3.4.15) imply that, if the set

204

Chapter 3. Control by the right hand sides in elliptic problems

3.5 Control of systems whose state is described by variational inequalities

3.5.1 Setting of the problem
So far we have considered optimal control problems when the state of the system is de ned as the solution of an elliptic problem, and this solution minimizes the quadratic functional in a corresponding Hilbert space. Now, we will investigate the optimal control problems in which the state of the system (the solution of the direct problem) is the solution of a variational inequality, i.e., the state of the system is the function minimizing a quadratic functional on some set from the initial space. Let H be a Hilbert space over R, and suppose that

a(w v) is a symmetric, continuous, coercive bilinear form on H H , e H is a convex, closed set in H:

We assume also that U is a re exive Banach space, U U ! H is compact.

(3.5.1) (3.5.2) (3:5:3)

H , the imbedding

For a given element g 2 U , the state of the system u(g) is de ned as the solution of the problem
e u(g) 2 H a(u(g) u(g)) ; 2(f + g u(g)) = infe a(v v) ; 2(f + g v)]

v 2H

(3.5.4)

By Theorem 3.1.1 and Remark 3.1.1, for every g 2 U there exists a unique function u(g) satisfying (3.5.4). Theorem 3.1.2 implies the function u(g) to be characterized by the formula
e u(g) 2 H

f being a xed element of H .

a(u(g) v ; u(g)) (f + g v ; u(g))

e v 2 H: 9

Further, suppose we are given a mapping P : H U ! R such that

v g ! P (v g) is a lower semicontinuous mapping of H U =

into R with respect to the topology of the product of weak topologies of H and U , and let a goal functional be of the form

(3:5:5)

g ! J (g) = P (u(g) g):

(3:5:6)

3.5. Control of systems whose state is described by variational inequalities 205

Here, the function u(g) is a solution of the problem (3.5.4). Suppose that

Uad U , Uad is bounded in the strong topology of U and

sequentially weakly closed.

(3:5:7) (3:5:8)

The optimal control problem consists in nding a function y such that

y 2 Uad

J (y) = g2U J (g): inf

ad

Theorem 3.5.1 Let H be a Hilbert space over R, let the conditions (3.5.1){(3.5.3) hold, and let the state of the system u(g) is de ned as the solution of the problem (3.5.4), f being a xed element of H , g 2 U . Assume that a goal functional J (g) is
determined by the relations (3.5.5) and (3.5.6), and the set Uad meets the condition (3.5.7). Then, the problem (3.5.8) has a solution.

3.5.2 The existence theorem

To prove Theorem 3.5.1, we need the following lemma.

Lemma 3.5.1 Let H be a Hilbert space over R, let the conditions (3.5.1){(3.5.3) be satis ed, and let a function f 2 H be given. Then, the function g ! u(g)
determined by the solution of the problem (3.5.4) is a continuous mapping of the space U equipped with the weak topology into the space H endowed with the weak topology.

Proof. 1) Let fgng be a sequence of elements of U such that gn ! g weakly in U (3:5:9) and let fu(gn)g be the corresponding sequence of solutions of the problem (3.5.4).
Let us show that there exists a positive number c such that

ku(gn)kH c
Introduce the notations

8n:

(3:5:10) (3.5.11) (3.5.12)

un = u(gn ) n (un ) = a(un un ) ; 2(f + gn un )

and assume the condition (3.5.10) is not valid. Then, from the sequence fung one can choose a subsequence fumg such that

kum kH ! 1:

(3:5:13)

206

Chapter 3. Control by the right hand sides in elliptic problems

By (3.5.1) and (3.5.12), we conclude that 8m (3:5:14) m (um ) c1 kum k2 ; 2kf + gm kH kum kH H where c1 is a positive number. By (3.5.3) and (3.5.9), there exists a positive number c2 such that kf + gmkH c2 8m: Combining this with (3.5.13) and (3.5.14), we get (3:5:15) m (um ) ! 1:
e Let now v be an arbitrary element of H . (3.5.1) and (3.5.9) yield ja(v v) ; 2(f + gm v)j c3 kvk2 + c4 kvkH = c5 8m H where c3 , c4 , and c5 are positive numbers. Hence, from (3.5.4) and (3.5.12), we obtain 8m: m (um ) c5 The latter inequality makes a contradiction with (3.5.15), so that (3.5.10) is true. 2) (3.5.3), (3.5.9), and (3.5.10) imply that from the sequence fgn un g one can choose a subsequence fgm umg such that gm ! g strongly in H (3.5.16) um ! w weakly in H: (3.5.17) e Since um 2 H , from (3.5.2) and (3.5.17) we deduce e w 2 H:

(3:5:18)
e v 2 H:

(3.5.4) and (3.5.11) yield

a(um um) ; 2(f + gm um) a(v v) ; 2(f + gm v)

Taking into account (3.5.16) and (3.5.17), we have lim inf a(um um) a(w w) m!1 lim (f + gm um) = (f + g w) m!1 lim (f + gm v) = (f + g v): m!1

(3:5:19)

(3.5.20)

By using (3.5.20), we pass in (3.5.19) to the limit in m for an arbitrary xed e element v from H . Then, we get e a(w w) ; 2(f + g w) a(v v) ; 2(f + g v) v 2 H: (3:5:21)

3.5. Control of systems whose state is described by variational inequalities 207

From this and (3.5.18), we see that w = u(g). We have only to prove that (3.5.17) holds true not only for the subsequence fumg, but for the whole sequence fung, i.e., un ! w weakly in H: (3:5:22) Indeed, assume that (3.5.22) is not valid. Then, there exist " > 0, z 2 H , and a subsequence fuk g of fun g such that

(3:5:24) Then, by passing to the limit just as it was done above, we infer that the function u0 meets the relation (3.5.21) in which w is replaced by u0 . For a given element e g 2 U , there exists a unique function w 2 H satisfying (3.5.21) (see Remark 3.1.1). Hence, u(g) = w = u0 and the relations (3.5.23) and (3.5.24) make a contradiction. So, (3.5.22) holds, and the lemma is proved. Proof of Theorem 3.5.1. Let fgng be a minimizing sequence, i.e.,

from it such that

j(z uk ; w)j " 8k: (3:5:23) Since the subsequence fuk g is bounded, we can choose a subsequence fulg
ul ! u0 weakly in H
e u0 2 H:

fgng 2 Uad

J (gn ) ! g2U J (g): inf

ad

(3:5:25)

By (3.5.7), we conclude that from the sequence fgng one can choose a subsequence fgmg such that

gm ! z weakly in U z 2 Uad :
By (3.5.26), using Lemma 3.5.1, we get

(3.5.26) (3.5.27) (3:5:28)

u(gm ) ! u(z )

weakly in H:

Further, the relations (3.5.5), (3.5.6), (3.5.26), and (3.5.28) yield lim inf J (gm ) J (z ): m!1 Combining this with (3.5.25) and (3.5.27), we have

J (z ) = g2U J (g): inf

ad

Thus, the function y = z is a solution to problem (3.5.8), concluding the proof.

208

Chapter 3. Control by the right hand sides in elliptic problems

3.5.3 An example of control of a system described by a variational inequality

Direct problem
Let H = H 1 ( ), where is a bounded open set in Rn with a smooth boundary S , and let a form a(w v) be determined by the relations (3.3.16){(3.3.18). Suppose e that the set H is of the form
e H = v j v 2 H 1 ( ) v 0 a.e. on S :

(3:5:29)

e In this case, H is a convex, closed cone in H with vertex at the origin. Indeed, e e the convexity of the set H is obvious, and the closedness of H in H is implied 2 by the continuity of the mapping v ! vjS of the space H 1 ( ) onto H 1 ( ) (see Theorem 1.6.5). Since the bilinear form a(w v) determined by the formulas (3.3.16){(3.3.18) is symmetric, continuous, and coercive on H 1 ( ) H 1 ( ), by virtue of Remark 3.1.4 for any element g 2 (H 1 ( )) there exists a unique function u(g) such that e u(g) 2 H

a(u(g) u(g)) ; 2(g u(g)) a(v v) ; 2(g v)

e u(g) 2 H

e v 2 H (3:5:30)

the function u(g) being characterized by the relations

a(u(g) v) (g v) a(u(g) u(g)) = (g u(g)): g(2) v ds

e v2H

(3.5.31) (3.5.32) (3:5:33)

Let an element g 2 (H 1 ( )) be determined by the relation (g v ) =

Z

g(1) v dx +

g(1) 2 L2 ( ) g(2) 2 L2 (S ):

Notice that the element g in (3.5.33) indeed determines a linear continuous form on H 1 ( ) (i.e., it belongs to (H 1 ( )) ) because v ! vjS is a continuous 1 mapping of H 1 ( ) onto H 2 (S ). The formulas (3.5.31) and (3.5.32) are interpreted in the following way. Let e v = ', where ' 2 D( ), so that v 2 H due to (3.5.29). Then, (3.5.31) implies

a(u(g) ') = (g ') Au(g) = ;

n X

' 2 D( ):
in (3:5:34)

From this, by using (3.3.16), and (3.5.33), we get

@ a @u(g) + a u(g) = g(1) ij @x 0 i i j =1 @xj

the derivatives being understood in the sense of distributions on .

3.5. Control of systems whose state is described by variational inequalities 209

e Multiplying (3.5.34) by v 2 H and applying Green's formula, we have Z Z ; @u(g) v ds + a(u(g) v) = g(1) v dx (3:5:35)

here

@A

see Lions and Magenes (1972), Aubin (1972). From (3.5.31), (3.5.33), and (3.5.35) Z @u(g) ; g(2) v ds 0: @
S A

n 1 @u(g) 2 H ; 2 (S ) @u(g) = X a @u(g) cos( x ) on S ij @x j @ A i j=1 @A i and cos( xj ) is the j -th component of the external unit normal to the boundary S of the domain . Note that, for a function u 2 H 1 ( ) such that1 Au = f 2 L2 ( ), one can determine uniquely @@u on S , in this case @@u 2 H ; 2 (S ), and Green's formula is A A valid: Z Z ; @@u v ds + a(u v) = fv dx v 2 H 1 ( ) S A

(3:5:36)

e Since here v 2 H , we have v 0 on S , so that the inequality (3.5.36) is equivalent to the following one @u(g) ; g(2) 0 on S (in the sense of H ; 1 (S )). 2 (3:5:37)

Moreover, by (3.5.32), (3.5.33), and (3.5.35) Z @u(g) ; g(2) u(g) ds = 0: Taking into account this equality, (3.5.37), and the relation u(g) 0 on S , we get u(g) @u(g) ; g(2) = 0 on S:
S

@A

@A

@A

(3:5:38) According to the last condition in (3.5.38), there exists a subset S0 of the g boundary S on which u(g) = 0 then @u(A) ; g(2) = 0 on S n S0 . It is obvious that @ S0 is not known beforehand, in particular, S0 may be an empty set.

Thus, in the space H 1 ( ), there exists a unique function u(g) that satis es the equation (3.5.34) and the following boundary conditions u(g) 0 on S @u(g) ; g(2) 0 on S u(g) @u(g) ; g(2) = 0 on S:

@A

@A

210

Chapter 3. Control by the right hand sides in elliptic problems

;

The optimal control problem

Let us consider the function g = g(1) g(2) 2 L2 ( ) L2 (S ) as a control. So,

U = L2 ( ) L2 (S ): (3:5:39) Through the formula (3.5.33), to every element g 2 U there corresponds a linear continuous functional in the space H 1 ( ), which will also be denoted g. Thus, we de ned the imbedding U ! (H 1 ( )) . 2 By Theorem 1.5.12, since the imbeddings H 1 ( ) ! L2 ( ) and H 1 (S ) ! 1 1 ( )) , L2 (S ) ! H ; 2 (S ). L2 (S ) are compact, so are the imbeddings L2 ( ) ! (H
the imbedding U ! (H 1 ( )) is compact. De ne a mapping P : H 1 ( ) ! R by Hence, (3:5:40) (3:5:41)

w ! P (w) =

n X Z bi i=1

@w ; z 2 dx + b Z (w ; z )2 dx 0 0 @xi i

zi and z0 being given functions from L2( ), bi being positive numbers.

Making use of Lemma 3.1.1, we see that the function w ! P (w) is lower semicontinuous in the weak topology of the space H 1 ( ). Let a goal functional J : U ! R have the form (3:5:42) (3:5:43)

g ! J (g) = P (u(g))

(3.5.31), (3.5.32)). Now, if the set Uad satis es the condition (3.5.7), then by (3.5.40), (3.5.42), and Theorem 3.5.1, we obtain that, for the goal functional J (g) determined by the relations (3.5.41) and (3.5.43), there exists an element y such that

u(g) being the solution of the problem (3.5.30) (or, equivalently, of the problem

y 2 Uad

J (y) = g2U J (g): inf

ad

Chapter 4

Direct problems for plates and shells

\You must next be told why a strong man came to fall a victim to a Beauty Hint" | O. Henry \The Indian Summer of Dry Valley Johnson"

4.1 Bending and free oscillations of thin plates

4.1.1 Basic relations of the theory of bending of thin plates

Consider a homogeneous thin plate of variable thickness (see Fig. 4.1.1). We suppose that the plate has a midplane such that the plate is symmetric with respect to it. Take the midplane of the plate to be the (x y) plane, and let the z axis be directed downwards. Denote displacements of points of the midplane along the z axis by w and assume that the so-called Kirchho hypotheses hold: 1. Normals to the midplane before the bending go over into the normals to the midsurface after the bending, and their length does not change during the bending. 2. Inside of the plate, the stresses normal to the midsurface are small as compared to other stress components, so that they can be neglected in relations between the stresses and strains (the plane stress state hypothesis). 3. Under bending, elements of the midsurface of the plate are not subject to tension and pressure. 211

212

Chapter 4. Direct problems for plates and shells

Figure 4.1.1: Thin plate of variable thickness Under these assumptions, the components u(x y z ) and v(x y z ) of the vector of displacements of points of the plate in the directions of the x and y axes have the form u(x y z ) = ;z @w v(x y z ) = ;z @w @x @y
2 2 "1 = @u = ;z @ w "2 = @v = ;z @ w @x @x2 @y @y2 2w @v @ "3 = @u + @x = ;2z @x@y : @y

where ; h z h and h is the thickness of the plate, depending on x and y. 2 2 The strain components are of the form

(4.1.1)

For an isotropic plate, the stress components are determined by the relations
1 2 3 2 2 E Ez = 1 ; 2 ("1 + "2 ) = ; 1 ; 2 @ w + @ w @x2 @y2 2 2 2 E Ez = 1 ; 2 ("2 + "1 ) = ; 1 ; 2 @ w + @ w @y2 @x2

(4.1.2)

@w = G"3 = 2(1E ) "3 = ; 1Ez @x@y + + where E is the elasticity modulus, G is the shearing modulus, and is the Poisson ratio. Introduce the notations
1 1 and 2 are called the components of the bending strains of the midplane, the component of the torsion strains.

= ;@ w @x2

= ;@ w @y2

@w = ; @x@y

(4:1:3)
3

is

4.1. Bending and free oscillations of thin plates

213

The bending moments M1 , M2 and the torque M3 are given by

M1 = M2 = M3 =

Z h 2

Z h 2 Z h 2

;h 2 ;h 2 ;h 2

1 (z ) z dz = D( 1 + 2 (z ) z dz = D( 2 + 3 (z ) z dz = D(1 ;

2) 1)

(4.1.4)

where D is the cylindrical sti ness of the plate, The energy of elastic deformation accumulated in the plate, i.e., the strain energy, is de ned by the following formula 1 ZZ dx dy Z 2 ( " + " + " ) dz (w) = 2 1 1 2 2 3 3 ;h 2
h

Eh3 D = 12(1 ; 2 ) :

(4:1:5)

(4:1:6)

where is the domain occupied by the midplane of the plate, which is supposed to be bounded in what follows. From this, using (4.1.1){(4.1.3), we obtain (w) = =
ZZ

D ( + )2 ; 2(1 ; )( 2 1 2
2

2 1 2 ; 3)

dx dy
2!

ZZ

@2w + @2w @x2 @y2

2 2 @2w ; 2(1 ; ) @ w @ w ; @x@y 2 @y 2 @x

dx dy:
(4.1.7)

4.1.2 Orthotropic plates

1 2 3

The strain components "1 , "2 , and "3 for an orthotropic plate are determined by the formulas (4.1.1), and the stress components are given by = E21 "1 + E22 "2 = ;E21 z @ w ; E22 z @ w @x2 @y2 = E11 "1 + E12 "2 = ;E11 z @ w ; E12 z @ w @x2 @y2
2 2 2 2

(4.1.8)

@w = G"3 = ;2Gz @x@y :

214 Here

Chapter 4. Direct problems for plates and shells

Eii = 1 ;Ei i=1 2 1 2 E12 = E21 = 2 E11 = 1 E22

(4.1.9)

E1 , E2 , G, 1 , and

2 being the elasticity characteristics of the material. By (4.1.2), (4.1.8), and (4.1.9), we conclude the isotropic material to be a partial case of the orthotropic one, and for it the following relations hold

E1 = E2 = E M1 = M2 = M3 =
Here,
1, 2,

1= 2=

G = 2(1E ) : +

For the bending moments and torque, we have in view of (4.1.8) and (4.1.9),
Z h 2 Z h 2 Z h 2
;h 2 ;h 2 ;h 2

1 (z ) z dz = D1 1 + D12 2 2 (z ) z dz = D21 1 + D2 2 3 (z ) z dz = D3 3 :

(4.1.10)

and

Taking into consideration the relations (4.1.1), (4.1.6), (4.1.8), and (4.1.10), we get the following expression for the strain energy of the orthotropic plate ZZ ; D1 2 + 2D12 1 2 + D2 2 + 2D3 2 dx dy (w ) = 1 1 2 3 2 (4.1.12) 1 Z Z (M + M + 2M ) dx dy: =2 1 1 2 2 3 3

3 i=1 2 Di = 12(1h;Ei ) 1 2 3 D12 = D21 = 2 D1 = 1 D2 D3 = h 6G :

are determined by the formulas (4.1.3) and (4.1.11)

4.1.3 Bilinear form corresponding to the strain energy of the plate

Isotropic plate
Let V be a closed subspace of W22 ( ) with the norm of the space W22 ( ). According to (4.1.7), to the strain energy of an isotropic plate there corresponds the following

4.1. Bending and free oscillations of thin plates

215

symmetric bilinear form on V

ah (u v) =

@2u + @2u @2v + @2v @x2 @y2 @x2 @y2 2 @2 @2 @2 @2u @2v ; (1 ; ) @xu @yv + @ u @xv ; 2 @x@y @x@y 2 2 @y 2 2 D ah (u u) = 2 (u) E = const > 0 h 2 Yp
0

ZZ

dx dy

u v 2 V:
(4.1.13)

Obviously,

where (u) is de ned by (4.1.7). Suppose that

(4:1:14) (4.1.15) (4.1.16) (4.1.17)

= const

<1

where the set Yp is de ned by the formulas (2.1.2) and (2.1.3). Then, (4.1.5) and (4.1.13) yield the continuity of the form ah in the following sense

jah (u v)j ckukV kvkV

Taking into account the inequality
2 2 2 @u@u @x2 @y2

u v 2 V h 2 Yp c = const > 0: @2u @x2

2 2 + @u @y2 2

(4:1:18) (4:1:19)

by (4.1.5), (4.1.13), and (4.1.17), we get

ah(u u)
=

ZZ

@2u @x2 @2u @x2

2 2

2 + @u @y2 2 + @u @y2

2 2

@2 2 @2u + 2 @xu @ u + 2(1 ; ) @x@y 2 @y 2 @u + @x@y

2 2

dx dy

c1

ZZ

dx dy
(4.1.20)

u 2 V h 2 Yp c1 = const > 0:
Now suppose that the following condition holds: the relations u 2 V , u = b0 + b1 x + b2y, where b0 , b1 , and b2 are constants, imply b0 = b1 = b2 = 0.

(4:1:21)

216

Chapter 4. Direct problems for plates and shells

Then, by Corollary 1.6.1, the formula

kuk1 =

ZZ

@2u @x2

2 + @u @y2

@u + @x@y

dx dy

1 2

(4:1:22)

de nes a norm in V , which is equivalent to the original one, i.e., to the norm of W22 ( ). Now, by (4.1.20) and (4.1.22), we have

ah (u u) c2 kuk2 V

u 2 V h 2 Yp c2 = const > 0:

(4:1:23)

Thus, the following theorem is proven.

ned by the relations (2.1.2) and (2.1.3). Also, let V be a closed subspace of W22 ( ) meeting (4.1.21). Then, the bilinear form ah (u v) determined by the formulas (4.1.5) and (4.1.13) is symmetric, continuous, and coercive on V V in the sense of the inequalities (4.1.18) and (4.1.23).

Theorem 4.1.1 Let the conditions (4.1.15){(4.1.17) hold and let a set Yp be de-

Orthotropic plate
ah ( u v ) =
ZZ

According to (4.1.3) and (4.1.12), to the strain energy of an orthotropic plate there corresponds the following bilinear form on V V
2 @2 2 @2 D1 @ u @xv + D2 @ u @yv 2 2 @x @y2 2

2 2 2 @2 @2u @2v +D12 @ u @ v + @ u @xv + 2D3 @x@y @x@y dx dy @x2 @y2 @y2 2 where Di , i = 1 2 3, and D12 are de ned by (4.1.11). Obviously, ah (u u) = 2 (u), where (u) is determined by the formulas (4.1.3) and (4.1.12).

(4.1.24)

Assume that

E1 E2 G are positive numbers, h 2 Yp 1 and 2 are constants, 0 i < 1 i = 1 2:

(4.1.25) (4.1.26) (4.1.27)

Making use of (4.1.11) and (4.1.25){(4.1.27), one can easily see that, for the bilinear form ah determined by (4.1.24), the inequality (4.1.18) is valid. From (4.1.24), by using (4.1.11), (4.1.19), and (4.1.25){(4.1.27), we obtain for arbitrary u 2 V , h 2 Yp ,

ah (u u)

ZZ

@2 D1 @xu 2

2 + D2 @ u @y2

@u + 2D3 @x@y

4.1. Bending and free oscillations of thin plates

217

; D12
c1
ZZ

@2u @x2
2

2 + @u @y2 2

2 2

dx dy
2

@2u @x2

2 + @u @y2

@u + @x@y

dx dy

c1 being a positive number.

Thus, we have proved the following

the sense of the inequalities (4.1.18) and (4.1.23).

ned by the relations (2.1.2) and (2.1.3). Assume V to be a closed subspace in W22 ( ) meeting (4.1.21). Then, the bilinear form ah (u v) determined by the formulas (4.1.11) and (4.1.24) is symmetric, continuous, and coercive on V V in

Theorem 4.1.2 Let the conditions (4.1.25){(4.1.27) hold and let a set Yp be de-

Remark 4.1.1 Theorem 4.1.2 remains valid if the conditions (4.1.27) are replaced by the following ones
1

and

are constants,

1 2<1

E1 ; 2 E2 > 0: 1

(4:1:28)

Indeed, by (4.1.25) and the Sylvester criterion, we deduce the quadratic form E1 2 + 2 1 E2 E2 2 1 2+ 1; 1 22R 1 1; 2 1; 1 2 1 2 1 2 to be positive de nite. Hence, we have

ah (u u) c

ZZ

@2u 2 + @2u 2 + @2u @x2 @y2 @x@y u 2 V h 2 Yp c = const > 0:

dx dy

In the sequel, in considering various orthotropic plates and shells, we suppose the conditions (4.1.27) to be ful lled, since the majority of actual orthotropic materials are of that kind. However, in what follows, the relations (4.1.27) could be replaced by less restrictive assumptions (4.1.28).

4.1.4 Problem of bending of a plate

Setting of the problem. Examples of boundary conditions and loads

Let f be a load acting on a plate, which is identi ed with some element of the space V . Then, the problem of the bending of the plate reduces to the problem of nding a function u such that

u2V

ah (u v) = (f v)

v2V

(4:1:29)

218

Chapter 4. Direct problems for plates and shells

the form ah being determined by the relation (4.1.13) for an isotropic plate and by the formula (4.1.24) for an orthotropic one. By Theorems 4.1.1 and 4.1.2, for any f 2 V and h 2 Yp , the problem (4.1.29) has a unique solution for the isotropic and orthotropic plates. From the physical point of view, the problem (4.1.29) means that one searches for a function u which minimizes the stored energy of the system on a set of functions meeting smoothness conditions and main boundary conditions which correspond to the fastening of the plate. The stored energy of the system (w) is determined by where ;(f w) is the potential energy of the load f . Let us consider some examples of the space V and the load f . Let S be the boundary of the domain occupied by the midplane of the plate, and S1 S (in particular, it may happen that S1 = S ). We suppose that S1 contains three points which do not belong to a single straight line, and the domain is bounded. De ne the space V as follows: (w) = 1 ah (w w) ; (f w) 2

w2V

(4:1:30)

V = u j u 2 W22 ( ) ujS1 = 0 :

(4:1:31)

In this case, the space V corresponds to the fastening (supporting) of the plate at the part S1 of the boundary. Let u = b0 + b1x + b2y, where b0 , b1 , b2 are constants and u 2 V . Since u vanishes in three points which do not belong to a single straight line, we easily deduce that b0 = b1 = b2 = 0, i.e., (4.1.21) holds. Let us consider the clamp of the plate at S1 . In this case, the space V looks as follows n o V = u j u 2 W22 ( ) ujS1 = 0 @u = 0 (4:1:32) @
@ @ being the derivative with respect to the normal to the boundary S . Here, by the theorem on the trace space (see Theorem 1.6.5 and Remarks 1.6.1, 1.6.2), the conditions ujS1 = 0 and @u S1 = 0 do, and (4.1.21) is valid. @ S1

In solving problems of plate bending, one uses also the condition of a simple support. In this case,
2 2 M = M1 x + M2 y + 2M3 x y = 0 on S1 x and y being the components of the exterior unit normal to S . This condition makes no sense for an arbitrary function u from W22 ( ) it is called the condition

u S1 = 0

of transversality, or the natural condition (see the Interpretation of the problem (4.1.29) below). Let now the space V be of the form

V = u j u 2 W22 ( ) u(x1 y1 ) = u(x2 y2 ) = u(x3 y3) = 0 :

(4:1:33)

4.1. Bending and free oscillations of thin plates

219

Here, (x1 y1 ), (x2 y2 ), and (x3 y3 ) are three points from which do not belong to a single straight line. Since the imbedding of W22 ( ) into C ( ) is continuous, the conditions u(xi yi ) = 0, i = 1 2 3, do, and the space V determined by (4.1.33) meets the condition (4.1.21). Notice that the space V de ned by (4.1.33) and similar ones appear in the calculation of elements of aircraft constructions when the plate is fastened at the center and at two parts of the edge surface by hinges (Fig. 4.1.2).
W
( x1 , y1 )

( x3 , y3 ) ( x2 , y2 )

Figure 4.1.2: Plate fastened at three points (x1 y1 ), (x2 y2 ), and (x3 y3 ) Now, instead of (4.1.33), one can use the following spaces

solution of the problem (4.1.29) was studied by Litvinov (1981a). All the spaces V de ned by the formulas (4.1.31){(4.1.35) are closed subspaces of W22 ( ). Let us consider an example of a load which is represented by an element f 2 V . Let the space V be determined by (4.1.31), let S1 be an open set in S , and let S2 be the interior of S n S1 . De ne an element f 2 V by

V = u j u 2 W22 ( ) u = 0 on 1 ujS2 = ujS3 = 0 (4.1.34) 2 ( ) u(x y ) = 0 uj = uj = 0 V = u j u 2 W2 (4.1.35) 3 3 S2 S3 (see Fig. 4.1.3), where 1 is a subset of , and S2 , S3 are segments. In the case when the space V is determined by the relation (4.1.35), an approximation of a

v ! (f v) =

ZZ

f1 v dx dy +

S2

f2 v ds +

@v f3 @ ds:

(4:1:36)

@v Here, @ is the derivative with respect to the normal of S , f1 is the distributed load, f2 is the cutting force, and f3 is the bending moment, which act on , S2 , and S , respectively. If we use the space V from (4.1.32), then the latter integral @v in (4.1.36) is to be taken over S2 , instead of S , because now @ = 0 on S1 .

220

Chapter 4. Direct problems for plates and shells

( x3 , y3 ) W
1

S2

S3

S2 , S3

Figure 4.1.3: Plate fastened at a subset We suppose also that

or at a point (x3 y3 ) and at segments

f1 2 L2( )

2 f2 2 H ; 3 (S2 )

2 f3 2 H ; 1 (S )

(4:1:37)

and the boundary S is regular enough. The results of subsec. 1.6.4 imply that, if (4.1.37) holds true, then the functional f de ned by (4.1.36) is continuous on V , i.e., f 2 V . Let us interpret the problem (4.1.29), supposing that the bilinear form ah is de ned by the relation (4.1.13), f by (4.1.36), and V by (4.1.31). Take in (4.1.29) v = ' 2 D( ). Then Au = f1 in (4:1:38) the operator A being de ned by the relation
2 2 2 @2 @2 @2 Au = @x2 D @ u + @ u + @y2 D @ u + @xu 2 @x2 @y2 @y2 2 2u @ @ + 2(1 ; ) @x@y D @x@y

Interpretation of the problem (4.1.29)

(4.1.39)

where the derivatives are taken in the sense of distributions on . Suppose that the function u is smooth. Then, by multiplying both hand sides of the equality (4.1.38) by v 2 V and making use of Green's formula, we get
ZZ

@v (Au)v dx dy = ; (B1 u) @ ds ;
S

S2

(B2 u)v ds + ah(u v)

4.1. Bending and free oscillations of thin plates

221 (4.1.40)

= Here,

ZZ

f1 v dx dy:

@2u + @2u + @x2 @y2 2 2 @2u 2 2 (1 ; ) @ u x + @ u y + 2 @x@y x y on S (4.1.41) @x2 @y2 2 2 @2u ; 2 2 B2 u = ; @@ D @ u + @ u ; (1 ; ) @@ D @x@y y ; x @x2 @y2 2u @2u @ @2u @2 + @x2 ; @y2 x y ; @D @x@y ; @D @xu y @x @y 2 2 u @D @ 2 u + @D @ 2 ; @y @x@y x on S2 : (4.1.42) @x @y Here, x and y are the components of the exterior unit normal to S , and @@ B1 u = D
is the tangent derivative, i.e., the derivative along the unit tangent vector (see Fig. 4.1.4). Notice that, if u 2 V is not smooth, but Au = f1 2 L2 ( ), then the relation (4.1.40) can be substantiated by using the abstract Green's formula (see Aubin (1972)). In this case,
1 B1 u 2 H ; 2 (S ) 3 B2 u 2 H ; 2 (S2 )

and the formula (4.1.40) takes the form

ZZ ZZ

@v (Au) v dx dy = ; B1 u @
=

; B2 u vjS2 + ah (u v)
v 2 V:
(4.1.43)

f1 v dx dy

Due to (4.1.29) and (4.1.36), we have

ah(u v) =

ZZ

; @v f1v dx dy + f2 vjS2 + f3 @ S

v 2 V:

(4:1:44)

This equality together with (4.1.43) implies

;

Now, Theorem 1.6.5 yields f2 = B2 u on S2

@v = 0: f2 ; B2 u vjS2 + f3 ; B1 u @ S f3 = B1 u on S:
(4:1:45)

222

Chapter 4. Direct problems for plates and shells

y

n
t

Figure 4.1.4: Domain

and unit normal and tangent vectors

Thus, the function u solving the problem (4.1.44) in which the form ah (u v) is de ned by (4.1.13) and fi meets (4.1.37) is shown to satisfy the conditions (4.1.38) and (4.1.45). Conversely, using Green's formula, we get
ZZ

@v (Aw) v dx dy = ; B1 w @

; B2 w vjS2 + ah (w v)

which is valid for any function w 2 V such that Aw 2 L2 ( ) and for any v 2 V (see Aubin (1972)), and one can easily see that, if the function u satis es (4.1.38) and (4.1.45), then it is a solution of the problem (4.1.44). By virtue of Theorem 4.1.1, the latter problem has a unique solution. We stress that the boundary conditions (4.1.45) are called the natural conditions, or the transversality conditions, and the boundary conditions de ning the space V (in this case u S1 = 0, see (4.1.31)) are called main or stable conditions. If in (4.1.44) f3 = 0, i.e., no bending moments act on the plate along the boundary, then one says that the plate has a simple support. By (4.1.45), we have the following natural boundary condition

B1 u = 0

on S:

(4:1:46)

If no bending moments and cutting forces act on the boundary S , i.e., f2 = 0 and f3 = 0 in (4.1.44), then we obtain the following natural boundary conditions

B1 u = 0 on S

B2 u = 0 on S2

which are referred to as the conditions of the free edge. In the same way, the problem (4.1.29) can be interpreted for the orthotropic plate.

4.1. Bending and free oscillations of thin plates

223

Let, as before, V be a closed subspace of W22 ( ) meeting (4.1.21). Denote by C 2 ((0 1) V ) the set of twice continuously di erentiable functions which are dened on the interval (0 1) and take values in V . Let the plate make free transversal oscillations. A function U (x y t) determining the de ection of the plate, which depends on the coordinates x, y and time t, is de ned to be the solution of the problem
ZZ
2 h@ U @t2

4.1.5 Problem of free oscillations of a plate

v + 12

h3

@3U @t2 @x

U 2 C 2 ((0 1) V ) @v + @ 3 U @v @x @t2 @y @y v 2 V t 2 (0 1):

(4.1.47)

dx dy + ah (U v) = 0
(4.1.48)

Here, the bilinear form ah is de ned by the relations (4.1.13) and (4.1.24) for isotropic and orthotropic plates, respectively, being the density of the material of the plate, = const > 0: (4:1:49) The rst term under the sign of integral in (4.1.48) determines the work of the inertia forces on virtual transversal displacements, and the second one determines the work of these forces on virtual longitudinal displacements. Indeed, by the Kirchho hypothesis (see subsec. 4.1.1), the components Px and Py of the inertia force distributed over the volume have the form

@3U Px = ; z @t2 @x

@3 Py = ; z @t2U : @y

If v is a virtual transversal displacement, i.e., a virtual de ection of the plate, then, by using again the Kirchho hypothesis, we conclude that the corresponding longitudinal displacements vx , vy in the x and y directions are

@v vx = ;z @x
ZZ

vy = ;z @v : @y dx dy
Z h 2
;h 2

The work of the inertia forces on virtual longitudinal displacements is given by

dx dy

Z h 2

;2

(P v + Py vy ) dz = h x x

ZZ

@ 3 U @v @ 3 z 2 @t2@x @x + @t2U @v dz: @y @y

Implementing the integration in z on the right hand side of this equality, we obtain the second term under the sign of integral in (4.1.48). We will search for the solution of the problem (4.1.47), (4.1.48) in the form

U (x y t) = (C1 cos !t + C2 sin !t)u(x y)

u(x y) 2 V

(4:1:50)

224

Chapter 4. Direct problems for plates and shells

C1 and C2 being constants.

The functions from (4.1.50) solving the problem (4.1.47), (4.1.48) are called natural oscillations of the plate. By substituting (4.1.50) into (4.1.48), we get the following equation for determining u(x y)

u2V bh (u v) =
ZZ

ah (u v) = bh (u v) v2V 2 =! h3 @u @v @v huv + 12 @x @x + @u @y dx dy: @y

(4.1.51) (4.1.52) (4.1.53)

The second term in (4.1.53) is called the work of forces caused by the inertia of rotation. For thin plates, h h3 , and so in many cases one can neglect the second term. Then, the form bh takes the form

bh(u v) =

ZZ

huv dx dy:

(4:1:54)

Theorem 4.1.3 Let the assumptions of Theorem 4.1.1 (for an isotropic plate ) or of Theorem 4.1.2 (for an orthotropic plate ) hold true and let (4.1.49) be ful lled. Let also the form bh be determined by the relation (4.1.53) or (4.1.54). Then, for any h 2 Yp , the spectral problem (4.1.51) for both the orthotropic and isotropic plates possesses a sequence of nonzero solutions fuig 2 V corresponding to a sequence of eigenvalues f i g such that
0<
and
1

ah (ui v) = i bh (ui v)
1 2 3

lim = 1 i!1 i

v2V

(4.1.55)

Each eigenvalue of the sequence f i g appears as many times as its multiplicity, and the multiplicity of each eigenvalue is nite.

(u u ) (u u) = ah(u 1 u 1) = uinf ah(u u) : bh 1 1 u2V bh 6=0

Proof. By virtue of Theorems 4.1.1 and 4.1.2 (for isotropic and orthotropic plates,

~ respectively), the bilinear form ah de nes a scalar product in V . Denote by V the space coinciding with V as a set and with the scalar product de ned by the form ~ ah . Making use of Theorems 4.1.1 and 4.1.2, it is easy to see that V is a Hilbert space. ~ Being considered as a topological space, V coincides with V and the norm generated by the form ah is equivalent to the one of the space V , i.e., to the norm of W22 ( ).

4.2. Problem of stability of a thin plate

225

The bilinear forms bh from (4.1.53) and (4.1.54) are obviously symmetric and continuous on W21 ( ). Moreover, for these forms, the following estimate holds

bh(v v) ckvk2 2( L
i=

v 2 W21 ( )

(4:1:56)

c being a positive constant. From (4.1.55) and (4.1.56) we conclude that ah (ui ui ) > 0 bh (ui ui )

8i:

To nish the proof, we use Theorem 1.5.8, taking into account that the imbedding ~ of V into W21 ( ) is compact, and pass from the problem (4.1.55) to the problem
i ah (ui v ) = bh (ui v )

v2V

where i = 1i .

4.2 Problem of stability of a thin plate

\The little hut was so wretched that it knew not on which side to fall, and therefore remained standing" | H. Ch. Andersen \The Ugly Duckling"

4.2.1 Stored energy of a plate

Let, as before, the (x y) plane coincide with the midplane of the plate, let the z axis be directed downwards (see Fig. 4.1.1), let be the domain occupied by the midplane of the plate, and let S be the boundary of . Suppose that the plate is fastened at a part S1 of the boundary, and at another part S2 = S n S1 the plate is exposed to longitudinal forces Q = ( Q1 Q2 ), which are proportional to a number parameter . Here, Q1 and Q2 are the components of the vector Q in the x and y axes, which are functions of s 2 S2 . Denote by u, v, w the components of the vector ! of displacements of points of the midplane in the x, y, z axes, i.e., ! = (u v w). The strain components of the midplane of the plate are determined by the following relations
2 @u @v 1 "11 (!) = @x + 1 @w "22 (!) = @y + 2 @w 2 @x @y @u + @v + @w @w "12 (!) = @y @x @x @y 2

(4.2.1)

226

Chapter 4. Direct problems for plates and shells

and the bending and torsion strains are described by the formulas (4.1.3). So, the strain components at a point (x y z ) of the plate, denoted by "z (!), "z (!), 11 22 "z (!), are given by 12

"z (!) = "11 (!) + z 1 (!) 11 "z (!) = "22 (!) + z 2 (!) 22 "z (!) = "12 (!) + 2z 3(!) 12
where

@ 2w 1 (! ) = ; @x2

@ 2w 2 (! ) = ; @y 2

@2w 3 (! ) = ; @x @y :

We assume that the material of the plate is orthotropic, and the principal directions of the elasticity of the material coincide with the directions of the x and y coordinate lines. Then, the stress components at a point (x y z ) of the plate are determined by
11 (! ) = E11 "11 (! ) + E12 "22 (! ) z z z 22 (! ) = E21 "11 (! ) + E22 "22 (! ) z z 12 (! ) = G"12 (! ):

Using the foregoing formulas, we get the following relation for the strain energy of the plate
Z 2 1 z z z =2 dx dy h 11 (!)"z (!) + 22 (!)"z (!) + 12 (!)"z (!) dz 11 22 12 ;2 ZZ c11 ("11 (!))2 + 2c12 "11 (!)"22 (!) + c22 ("22 (!))2 + c33 ("12 (!))2 ] dx dy =1 2 1 ZZ D ( (!))2 + 2D (!) (!) + D ( (!))2 + 2D ( (!))2 ] dx dy: +2 1 1 12 1 2 2 2 3 3 (4.2.2)
h

1 (! ) = ZZ

The rst term on the right hand side of (4.2.2) is the energy caused by the deformations of the midplane of the plate, while the second one coincides with the expression (4.1.12) and determines the bending and torsion energies. The coe cients in (4.2.2) are given by the following formulas

Di = 12 Eii

h3

cik = hEik i=1 2

i k=1 2

c33 = hG

D12 = 2 D1 = 1 D2

3 D3 = h 6G

4.2. Problem of stability of a thin plate

227 (4.2.3) (4:2:4)

In particular, if

Eii = 1 ;Ei

1 2

i=1 2

E12 = E21 = 2 E11 = 1 E22 :

1= 2=

E1 = E2 = E

G = 2(1E ) +

then the relations (4.2.2) and (4.2.3) determine the energy of elastic deformation accumulated in the isotropic plate. The stored energy of the orthotropic plate is (!) =
1 (! ) ;

The functions Q1 and Q2 can vanish at a part of S2 . At the part S1 = S n S2 , the plate is fastened. Thus, we will consider the functional (4.2.5) in the space V given by V = V1 V2 (4:2:6) and if ! = (u v w) 2 V , then (u v) 2 V1 and w 2 V2 . The space V1 is de ned by V1 = g j g = (g1 g2 ) g 2 (W21 ( ))2 g jS1 = 0 (4:2:7) and equipped with the norm of the space (W21 ( ))2 , i.e., kgkV1 = kgk(W21( ))2 g 2 V1 : (4:2:8) In this section, the set S1 is assumed to be of positive one-dimensional measure. From the results of subsec. 1.7.2, we conclude that the formula

S2

(Q1 u + Q2 v) ds:

(4:2:5)

kgk1 =

ZZ

@g1 @x

+ @g2 @y

+ @g1 + @g2 @y @x

dx dy

1=2

(4:2:9)

de nes a norm in the space V1 which is equivalent to the original one, i.e., to the norm of (W21 ( ))2 . In other words, there exist positive numbers c1 and c2 such that c1 kgk1 kgkV1 c2 kgk1 g 2 V1 : (4:2:10) We assume the space V2 to meet the following condition

V2 is a closed subspace of W22 ( ), the conditions q 2 V2 ,= (4:2:11) q = b0 + b1 x + b2 y, where b0 , b1 , and b2 are constants, imply that b0 = b1 = b2 = 0. The space V2 is endowed with the norm of W22 ( ), so that we denote kqkV2 = kqkW22 ( ) q 2 V2 : (4:2:12)

228

Chapter 4. Direct problems for plates and shells

In particular, if the plate is supported on S1 , then

V2 = q j q 2 W22 ( ) qjS1 = 0 and if it is clamped on S1 , then @q = 0 V2 = q j q 2 W22 ( ) qjS1 = 0 @ S1

@ @ being the derivative with respect to the normal to S .

(4:2:13) and the relations (4.1.25){(4.1.27) hold. Then, the functional (4.2.5) is well-de ned on the whole space V . We will search for the stationary points of the functional (4.2.5) in the space V , which characterize the equilibrium states of the plate being under the longitudinal load Q. In other words, we look for functions ! such that

Suppose that

Q = (Q1 Q2 ) 2 (L2 (S2 ))2

! 2V

d dt

(! + th) t=0 = 0

h 2 V:

(4:2:14)

4.2.2 Conditions of stationarity

De ne bilinear continuous mappings as follows

T1 T2 T12 2 L2 (V2 V2 L2 ( )) w0 w00 2 V2 0 00 0 00 T1 (w0 w00 ) = 1 E11 h @w @w + 2 @w @w 2 @x @x @y @y 0 @w00 0 00 1 T2 (w0 w00 ) = 2 E22 h @w @y + 1 @w @w @y @x @x 0 @w00 0 @w00 1 T12 (w0 w00 ) = 2 Gh @w @y + @w @x : @x @y
Further, introduce linear continuous mappings

(4.2.15)

(4.2.16)

N1 N2 N12 2 L(V1 L2 ( ))
by the relations

(4:2:17)
2

g = (g1 g2) 2 V1

N1 (g) = E11 h @g1 + @x

@g2 @y

4.2. Problem of stability of a thin plate

229

midplane are determined by formulas ~ N1(g w) = N1 (g) + T1 (w w) ~ N2(g w) = N2 (g) + T2 (w w) ~ N12 (g w) = N12 (g) + T12(w w): From (4.2.15) and (4.2.17), we conclude ~ Ni : V1 V2 ! L2 ( ) i=1 2 ~12 : V1 V2 ! L2 ( ): N Also, introduce a bilinear continuous mapping N 2 L2 (V2 V2 V1 ) through the following relations w0 w00 2 V2 g = (g1 g2) 2 V1 (N (w0 w00 ) g) = ;
ZZ

! = (0 0 w), w 2 V2 , and the relation (4.2.18) de ne the forces in the midplane which are caused by the vector of displacements ! = (g1 g2 0), g = (g1 g2 ) 2 V1 . In the general case, when ! = (g1 g2 w) 2 V1 V2 , the forces acting in the
(4.2.19)

acting in the midplane of the plate which are caused by the vector of displacements

N12 (g) = Gh @g1 + @g2 : (4.2.18) N2 (g) = E22 h @g2 + 1 @g1 @y @x @y @x We point out that T1 (w w), T2 (w w), and T12 (w w) determine inner forces

(4.2.20) (4.2.21)

T1(w0 w00 ) @g1 + T2 (w0 w00 ) @g2 @x @y +T12 (w0 w00 ) @g1 + @g2 dx dy: @y @x

(4.2.22) (4:2:23)

(4.2.16) and (4.2.22) yield N to be a symmetric mapping, i.e., N (w0 w00 ) = N (w00 w0 ) w0 w00 2 V2 : Further, de ne a mapping A : V1 V2 ! V2 by the formulas g = (g1 g2 ) 2 V1 w0 w00 2 V2 A(g w0 ) 2 V2 ZZ 00 0 00 0 ~ ~ (A(g w0 ) w00 ) = N1 (g w0 ) @w @w + N2 (g w0 ) @w @w

@y @y 0 00 0 00 ~ + N12 (g w0 ) @w @w + @w @w @x @y @y @x

@x @x

dx dy:
(4.2.24)

230

Chapter 4. Direct problems for plates and shells

Let ! be a function satisfying the condition (4.2.14), i.e., it is a stationery point of the functional . Introduce the notations

! = (u v w ) g = (u v ):
It is easy to see that (4.2.14) is equivalent to the following system

(4.2.25) (4.2.26) (4.2.27) (4.2.28)

a(1) (g g) = (N (w w ) g) + (Q g) g 2 V1 (2) (w w) + (A(g w ) w) = 0 a w 2 V2 :

Here, we used the notations

a(1) (p g) =

ZZ

ZZ

h E11 @p1 @g1 + E22 @p2 @g2 + E12 @p1 @g2 + @p2 @g1 @x @x @y @y @x @y @y @x + G @p1 + @p2 @g1 + @g2 dx dy @y @x @y @x N1 (p) @g1 + N2 (p) @g2 + N12 (p) @g1 + @g2 dx dy @x @y @y @x p = (p1 p2 ) 2 V1 g = (g1 g2 ) 2 V1 Z (Q g) = (Q1 g1 + Q2g2 ) ds
(4.2.29) (4.2.30)

a(2) (w0 w00 ) =

+ E12

h3 E @ 2 w0 @ 2 w00 + E @ 2 w0 @ 2 w00 22 @y 2 @y 2 12 11 @x2 @x2 @ 2 w0 @ 2 w00 + @ 2 w0 @ 2 w00 + 4G @ 2 w0 @ 2 w00 dx dy @x2 @y2 @y2 @x2 @x@y @x@y 0 00 w w 2 V2 :

ZZ

S2

(4.2.31)

Notice that the equality (4.2.29) follows from (4.2.3) and (4.2.18).

Lemma 4.2.1 Let the conditions (4.1.25){(4.1.27) be ful lled and let a set Yp be de ned by the formulas (2.1.2) and (2.1.3). Further, let a space V1 be determined by ; (4.2.7) and endowed with the norm of W21 ( ) 2 (see (4.2.8)). Then, the bilinear form a(1) de ned by (4.2.3) and (4.2.29), is symmetric, continuous, and coercive on V1 V1 for every h 2 Yp , i.e.,
a(1) (p g) = a(1) (g p) p g 2 V1
(4.2.32)

4.2.3 Auxiliary statements

4.2. Problem of stability of a thin plate

231

ja(1) (p g)j c1 kpkV1 kgkV1 ~ p g 2 V1 (4.2.33) a(1) (g g) c2 kgk2 1 ~ V g 2 V1 (4.2.34) c1 and c2 being positive numbers that are independent of h 2 Yp . ~ ~ Proof. The relations (4.2.32) and (4.2.33) are obvious. Let us prove the inequality
(4.2.34). From the formulas (2.1.2), (2.1.3), (4.2.3), and (4.2.29), we deduce that

a(1) (g

g)

ZZ

h E11 @g1 @x

+ E22 @g2 @y

; 2 1 E22 @g1 @g2 @x @y

2

+ G @g1 + @g2 @y @x From this, taking into account the estimate

dx dy

g 2 V1 : (4.2.35)

@g1 2 + @g2 2 2 @g1 @g2 @x @y @x @y (4.1.27), and the relation 1 E22 = 2 E11 (see (4.2.3)), we get a(1) (g g) kkgk2 g 2 V1 : (4:2:36) 1 Here, k is a positive number independent of h 2 Yp , and kgk1 is determined by (4.2.9). Now, (4.2.34) is a consequence of (4.2.10) and (4.2.36), concluding the proof. By virtue of (4.2.21), Lemma 4.2.1, and the Riesz theorem, we have that, for any w 2 V2 , there exists a unique function P (w) such that P (w) 2 V1 a(1) (P (w) g) = (N (w w) g) g 2 V1 : (4:2:37) Thus, we have de ned the mapping V2 3 w ! P (w) 2 V1 . Lemma 4.2.2 Let the hypotheses of Lemma 4.2.1 be ful lled, let a space V2 be de ned by (4.2.11) and equipped with the norm of W22 ( ) (see (4.2.12)). Let also a mapping N be determined by the relations (4.2.16) and (4.2.22). Then, P : w ! P (w), P (w) being de ned by (4.2.37), is a continuously Frechet di erentiable mapping of V2 into V1 , i.e., at every point w 2 V2 there exists the Frechet 0 0 derivative Pw 2 L(V2 V1 ) of the mapping P and, moreover, w ! Pw is a continuous mapping of V2 into L(V2 V1 ). Proof. Let w and h be arbitrary elements from V2 . Taking into consideration (4.2.21) and (4.2.23), from (4.2.37) we get that a(1) (P (w + h) g) = (N (w + h w + h) g) = (N (w w) g) + 2(N (w h) g) + (N (h h) g) g 2 V1 : (4.2.38)

232

Chapter 4. Direct problems for plates and shells

Thus, (4.2.37) implies that a(1) (P (w + h) g) = a(1) (P (w) g) + 2(N (w h) g) + (N (h h) g)

g 2 V1 :

(4.2.39)

The operator R is de ned by the relation ~ ~ a(1) (Rf q) = (f q) f 2 V1 q 2 V1 : By (4.2.15) and (4.2.22), we conclude ~ j(N (h h) g)j ckhk2 2 kgkV1 h 2 V2 g 2 V1 c = const > 0: ~ V This yields kN (h h)kV1 ckhk2 2 h 2 V2 : ~ V Further, taking into account (4.2.34), (4.2.41), and (4.2.44), we obtain kR N (h h)kV1 c1 kR N (h h)kV1 c2 khk2 2 h 2 V2 ~ V c1 and c2 being positive numbers. From this inequality we deduce that kR N (h h)kV1 ! 0 as khk ! 0:

By virtue of Lemma 4.2.1, the bilinear form a(1) generates a scalar product and a norm on V1 , which is equivalent to the original norm of the space V1 , i.e., ; ~ the norm of W21 ( ) 2 . Hence, one can consider the Hilbert space V1 as the set of elements of V1 equipped with the scalar product and norm generated by the ~ bilinear form a(1) . (Obviously, V1 and V1 coincide as topological spaces, since they contain the same elements and their norms are equivalent.) Then, by (4.2.39) we have P (w + h) = P (w) + 2R N (w h) + R N (h h): (4:2:40) Here, R is the Riesz operator (see subsec. 1.5.1) ~ ~ R 2 L(V1 V1 ) kRkL(V1 V1 ) = 1: (4:2:41) ~ ~ (4:2:42) (4.2.43) (4:2:44)

(4.2.21) and (4.2.41) imply that, for a given w 2 V2 , the function h ! 2R N (w h) is a linear continuous mapping of V2 into V1 . Now, (4.2.40) and (4.2.45) yield that, at every point w 2 V2 , the mapping P : V2 ! V1 is Frechet di erentiable (see subsec. 1.9.1) and the Frechet derivative 0 Pw 2 L(V2 V1 ) is given by the formula 0 Pw h = 2R N (w h): (4:2:46)

khkV2

V2

(4:2:45)

4.2. Problem of stability of a thin plate

233 (4:2:47)

Let fwn g be a sequence of elements of V2 such that wn ! w0 in V2 : By using (4.2.21), (4.2.41), and (4.2.46), we have ; 0 0 Pwn ; Pw0 h V1 = 2kR N (wn h) ; R N (w0 h)kV1 = 2kR N ((wn ; w0 ) h)kV1

ckwn ; w0 kV2 khkV2 h 2 V2 c = const > 0:

L(V2 V1 )

From this, we get

0 0 Pwn ; Pw0

ckwn ; w0 kV2 :

At last, (4.2.47) implies

(Notice that, in this section, the form ah de ned by the formulas (4.1.11) and (4.1.24) is denoted by a(2) and the space V2 corresponds to the space V introduced in Section 4.1.) ~ The spaces V2 and V2 obviously coincide provided they are considered as topological ones, since they contain the same elements and are equipped with equivalent norms. Now, the function w ! P (w), P (w) determined by (4.2.37), ~ can be treated as a mapping of V2 into V1 , and Lemma 4.2.2 implies the following ~ Corollary 4.2.1 Let the hypotheses of Lemma 4.2.2 are ful lled and let V2 be the Hilbert space containing the elements of V2 , with the norm and scalar product de ned by the formulas (4.2.31) and (4.2.48). Then, the function w ! P (w), P (w) ~ determined by (4.2.37), is a continuously Frechet di erentiable mapping of V2 into V1 .

concluding the proof. Theorem 4.1.2 yields that the bilinear form a(2) (w0 w00 ) determined by the relation (4.2.31), is symmetric, continuous, and coercive on V2 V2 , and one can ~ de ne a Hilbert space V2 as the set of functions from V2 endowed with the norm and scalar product which are generated by the form a(2) , i.e., (w0 w00 )V2 = a(2) (w0 w00 ) kwk2~2 = a(2) (w w): (4:2:48) ~ V

0 0 Pwn ! Pw0 in L(V2 V1 )

4.2.4 Transformation of the problem (4.2.27), (4.2.28)

From (4.2.13), (4.2.30), Lemma 4.2.1, and the Riesz theorem, we conclude the existence of a unique function g such that ~ g 2 V1 ~ a(1) (~ g) = (Q g) g g 2 V1 : (4:2:49)

234

Chapter 4. Direct problems for plates and shells

Introduce the bilinear form ~ ~ bg 2 L2 (V2 V2 R) ~ given by (4:2:50)

bg (w0 w00 ) = ~

ZZ

0 00 0 00 N1 (~) @w @w + N2 (~) @w @w g @x @x g @y @y 0 00 0 00 + N12 (~) @w @w + @w @w g @x @y @y @x dx dy

~ w0 w00 2 V2 (4.2.51)

where N1 (~), N2 (~), and N12 (~) are given by the formulas (4.2.18). g g g By (4.2.27), (4.2.37), and (4.2.49), we get

g = g + P (w ): ~

(4:2:52)

Substituting this equality into (4.2.28) and taking into consideration (4.2.19), (4.2.24), and (4.2.51), we reduce the problem (4.2.27), (4.2.28) to that of nding a function w such that ~ w 2 V2 (2) (w w) + b (w w) + (Gw w) = 0 ~ a w 2 V2 : (4.2.53) g ~ ~ ~ Here, G is the mapping of V2 into V2 determined by (Gw w ) =
0 00 0 00 N1 (P (w0 )) + T1(w0 w0 )] @w @w @x @x 00 0 + N2 (P (w0 )) + T2 (w0 w0 )] @w @w @y @y 0 00 0 00 + N12 (P (w0 )) + T12 (w0 w0 )] @w @w + @w @w @x @y @y @x

ZZ

dx dy
(4.2.54)

~ w w 2 V2 :
0 00

If found is a function w solving the problem (4.2.53), then one can determine the function g from the equation (4.2.52). We will need now some properties of the mapping G.

Lemma 4.2.3 Let the conditions (4.1.25){(4.1.27) be ful lled and let a set Yp be determined by the relations (2.1.2), (2.1.3). Assume that spaces V1 and V2 are de~ ned by the expressions (4.2.7), (4.2.8), (4.2.11), and (4.2.12), and V2 is the set of ~ ~ functions from V2 equipped with the norm (4.2.48). Then, the mapping G : V2 ! V2

4.2. Problem of stability of a thin plate

235

given by the relations (4.2.16), (4.2.18), (4.2.22), (4.2.37), and (4.2.54) is contin~ uously Frechet di erentiable, i.e., at every point w 2 V2 there exists the Frechet derivative G0w of the mapping G and, moreover, w ! G0w is a continuous mapping ~ ~ ~ of V2 into L(V2 V2 ) and the equality G0 = holds, where is the zero element ~2 , and is the zero element of the space L(V2 V2 ). ~ ~ of the space V

Proof. Let us represent the mapping G in the form

(Gw0 w00 ) =
4 X

i=1

(Gi w0 w00 )

~ w0 w00 2 V2

(4:2:55)

where (G1 w w ) =
0 00

ZZ ZZ ZZ ZZ

(G2 w0 w00 ) = (G3 w0 w00 ) = (G4 w w ) =

0 00

0 00 N1 (P (w0 )) @w @w dx dy @x @x 0 00 N2 (P (w0 )) @w @w dx dy @y @y 00 0 00 0 N12 (P (w0 )) @w @w + @w @w dx dy @x @y @y @x 0 00 0 00 T1 (w0 w0 ) @w @w + T2(w0 w0 ) @w @w @x @x @y @y 00 0 00 0 + T12 (w0 w0 ) @w @w + @w @w @x @y @y @x dx dy:

(4.2.56)

~ It is su cient to verify that Lemma 4.2.3 is valid for every mapping Gi : V2 ! ~2 , i = 1 2 3 4. The mapping G1 is the composition of the mappings A and B , V i.e., G1 = B A: Here ~ A : V2 ! L2 ( ) W21 ( ) w ! Aw = N1 (P (w)) @w @x ~ and B is a bilinear, continuous mapping of the space L2 ( ) W21 ( ) into V2 if (p g) 2 L2 ( ) W21 ( ), then the value of the functional B (p g) at an element ~ w 2 V2 is given by the formula (B (p g) w) =
ZZ

pg @w dx dy: @x

From Lemma 4.2.2 and Property 4 of the Frechet derivative (see subsec. 1.9.1), we deduce that the mapping A is Frechet di erentiable at every point

236

Chapter 4. Direct problems for plates and shells

~ w 2 V2 and the derivative A0w has the form 0 0 A0w w0 = N1 Pw w0 @w @x

ZZ

~ w0 2 V2 :

Further, the mapping B is also Frechet di erentiable, and the derivative B(0p g) is given by (B(p g) (p1 g1) w) =
0

(p1 g + pg1 ) @w dx dy @x

~ (p g) (p1 g1) 2 L2 ( ) W21 ( ) w 2 V2 : From Property 4 of the Frechet derivative, we conclude the mapping G1 to ~ be Frechet di erentiable at any point w 2 V2 and the derivative G01w is given by (G1w w w ) =
0 0 00

ZZ

0 00 0 (N1 Pw w0 ) @w + (N1 P (w)) @w @w dx dy @x @x @x

~ w0 w00 2 V2 :

(4.2.57)

Taking into account that w ! P (w) is a continuously di erentiable mapping ~ of V2 into V1 (see Lemma 4.2.2), and using (4.2.17) and (4.2.57), it is easy to see ~ ~ ~ that w ! G01w is a continuous mapping of V2 into L(V2 V2 ). ~, where and ~ are the zero elements of the (4.2.37) implies that P ( ) = ~ spaces V2 and V1 , respectively. Thus, (4.2.57) yields

G01 =
~ ~ being the zero element of the space L(V2 V2 ). So, Lemma 4.2.3 is proved to be valid for the mapping G1 . In the same way, we derive it to hold true for the mappings G2 and G3 . At last, G4 is a trilinear ~ ~ continuous mapping of V2 into V2 , and the validity of Lemma 4.2.3 for it is also easily veri ed. A nonlinear operator A acting from a Hilbert space H into its dual space H is said to be compact if the image A(Q) of every bounded set Q H is relatively compact, i.e., the closure A(Q) of the set A(Q) is a compactum in H . Obviously, if the condition

fung H un ! u0 weakly in H
implies Aun ! Au0 strongly in H , then the operator A is compact. ~ ~ of V2 into V2 .

Lemma 4.2.4 Under the suppositions of Lemma 4.2.3, G is a compact mapping

4.2. Problem of stability of a thin plate

237

~ Proof. Let fwng be a sequence of elements of V2 such that ~ wn ! w0 weakly in V2 : Hence, due to the compactness of the imbedding W22 ( ) into W41 ( ) (see Theorem 1.6.2), we have wn ! w0 strongly in W41 ( ): (4:2:58) By (4.2.16), (4.2.22), and (4.2.58), we get N (wn wn ) ! N (w0 w0 ) strongly in V1 : (4:2:59) Now, taking into account Lemma 4.2.1, (4.2.37), and (4.2.59), we obtain P (wn ) ! P (w0 ) strongly in V1 : (4:2:60) Further, by virtue of (4.2.16), (4.2.17), (4.2.54), (4.2.58), and (4.2.60), we deduce ~ Gwn ! Gw0 strongly in V2 concluding the proof. ~ Consider the following eigenvalue problem: Find ( w) 2 R V2 , w 6= such that ~ a(2) (w w0 ) + bg (w w0 ) = 0 w0 2 V2 : (4:2:61) ~ Since the bilinear form bg is symmetric (see (4.2.51)), the imbedding of W22 ( ) ~ ~ into W41 ( ) is compact and the bilinear form a(2) de nes a scalar product in V2 , from Theorem 1.5.8 we obtain the following Lemma 4.2.5 Let the conditions (4.1.25){(4.1.27) be ful lled and let a set Yp be ~ determined by the relations (2.1.2), (2.1.3). Let V2 be the Hilbert space containing the elements of the space V2 (see (4.2.11)) with the scalar product generated by the form a(2) , and let bg be the bilinear form de ned by the formulas (4.2.30), (4.2.49), ~ (4.2.51), and let (4.2.13) be valid. Then, the problem (4.2.61) has a countable set of eigenvalues f i g 0 < j 1j j 2j nlim j n j = 1: !1
Each eigenvalue appears as many times as its multiplicity, and the multiplicity of each eigenvalue is nite.

We will need now the notion of a bifurcation. Let H be a Hilbert space over R, let A be a continuous mapping of H into its dual space H , and let B be a compact mapping of H into H . Let also A =B =

4.2.5 Stability of a plate and bifurcation

238

Chapter 4. Direct problems for plates and shells

where and are the zero elements of the spaces H and H , respectively. A real number 0 is said to be a bifurcation point of the equation

Au + Bu =
if for any " > 0 there exist u" and such that 0 < ku" kH < ":

(4:2:62)

Au" + Bu" =

j 0; j<

Let us again consider the problem (4.2.53). The solution of this problem determines the transversal displacements of the plate (the function w ) on which the longitudinal forces Q act, the forces being distributed on the part S2 of the boundary S (see (4.2.30)). By virtue of (4.2.54), G = 1 , where and 1 are the zero elements of ~ ~ the spaces V2 and V2 , respectively. Thus, for each 2 R, there exists the trivial solution w = of the problem (4.2.53). We are interested in the bifurcation points of the problem (4.2.53), i.e., the values of in neighborhoods of which nontrivial solutions of the problem (4.2.53) appear. It is natural to connect with the smallest positive bifurcation point the loss of the stability of the plate. We note that the problem (4.2.53) reduces to the problem (4.2.62) if one ~ ~ ~ ~ de nes the operators A : V2 ! V2 and B : V2 ! V2 by (Aw0 w00 ) = a(2) (w0 w00 ) + (Gw0 w00 ) (Bw0 w00 ) = bg (w0 w00 ) ~ ~ where w0 w00 2 V2 .

Proof. By the Riesz theorem, the bilinear form bg generates an operator Bg 2 ~ ~ ~ ~ L(V2 V2 ) such that ~ bg (w0 w00 ) = a(2) (Bg w0 w00 ) w0 w00 2 V2 : (4:2:64) ~ ~
Since the imbedding of W22 ( ) into W41 ( ) is compact, ~ ~ Bg is a linear compact mapping of V2 into V2 . ~ (4:2:65)

Theorem 4.2.1 Let the conditions (4.1.25){(4.1.27) be ful lled and let a set Yp ~ be de ned by the relations (2.1.2) and (2.1.3). Let also V2 be the Hilbert space containing the elements of the space V2 (see (4.2.11)) equipped with the scalar product generated by the form a(2) , let bg be the bilinear form determined by the formulas ~ ~ ~ (4.2.30), (4.2.49), and (4.2.51), let (4.2.13) hold, and let a mapping G : V2 ! V2 be de ned by (4.2.54). If 0 is the bifurcation point of the problem (4.2.53), then 0 is an eigenvalue of the problem (4.2.61), i.e., there exists a function w0 such that ~ ~ w0 2 V2 w0 6= a(2) (w0 w) + 0 bg (w0 w) = 0 w 2 V2: (4:2:63) ~

4.2. Problem of stability of a thin plate

239

~ Further, let R1 be the Riesz operator in the space V2 which is de ned by the relation ~ ~ (f w) = a(2) (R1 f w) f 2 V2 w 2 V2 : (4:2:66) Taking into account (4.2.64) and (4.2.66), the problem (4.2.53) can be represented in the form ~ w 2 V2 w + Bg w + R1 Gw = : (4:2:67) ~ The problem (4.2.63) now takes the form ~ w0 2 V2 w0 6= w0 + 0 Bg w0 = : (4:2:68) ~ ~ ~ De ne a mapping P2 : R V2 ! V2 by

P2 ( w) = w + Bg w + R1 Gw ~

~ ( w) 2 R V2 :

(4:2:69) (4:2:70)

Then, the problem (4.2.67) reduces to the following one ~ w 2 V2 P2 ( w ) = :

Making use of Lemma 4.2.3 and of the properties of Frechet derivative (see subsec. 1.9.1), one easily sees that, for every xed 2 R, the function w ! P2 ( w) has the Frechet derivative which is given by the formula

@P2 ( w) 2 L(V V ) ~2 ~2 @w

(4:2:71)

@P2 ( w) = I + B + R G0 : (4:2:72) g ~ 1 w @w ~ Here, I is the identity operator in V2 . Since G0 = (see Lemma 4.2.3), from
) = I + Bg : (4:2:73) ~ ~ ~ ~ Since w ! G0w is a continuous mapping of V2 into L(V2 V2 ) (see Lemma ~ ~ 4.2.3) and R1 2 L(V2 V2 ), from (4.2.72) we deduce that
) ~ w ! @P2 ( w) is a continuous mapping of R V2 into @w ~ ~ L(V2 V2 ).

(4.2.72) we conclude

@P2 ( @w

(4:2:74)

Let 0 be a bifurcation point of the problem (4.2.53). Then, there exists a sequence f n wn g such that j ; j< 1 0 < kw k < 1 (4.2.75)
n
0

~ n V2

240

Chapter 4. Direct problems for plates and shells

P2 ( n wn ) = : P2 (
By virtue of (4.2.73), we have )=

(4.2.76) (4:2:77) (4:2:78)

For any 2 R, the function w = is a solution to the problem (4.2.53), so that

2 R:

@P2 ( @w

) = I + 0 Bg : ~

Then, by virtue of (4.2.74) and (4.2.79), from the implicit function theorem (Theorem 1.9.1) we deduce the existence of an open set U1 in R containing the ~ point 0 , and of an open set U2 in V2 containing the point w = such that, for all 2 U1 , the equation P2 ( w) = has a unique solution with respect to w. Hence, by (4.2.77), this solution is w = for all 2 U1 . Thus, the relations (4.2.75) and (4.2.76) make a contradiction, concluding the proof. So, only the eigenvalues of the problem (4.2.61) can be bifurcation points of the problem (4.2.53). The statement below gives a su cient condition of the existence of bifurcation points.

If 0 is not an eigenvalue of the problem (4.2.61), or, equivalently, of the problem (4.2.68), then (4.2.65) together with the Fredholm alternative implies that @P2 ( ) is an invertible mapping of V into V . ~2 ~2 (4:2:79) @w 0

Theorem 4.2.2 Under the suppositions of Theorem 4.2.1, each eigenvalue of

the problem (4.2.61) having odd multiplicity is a bifurcation point of the problem (4.2.53). ~ ~ Proof. Introduce a mapping : V2 ! V2 by

= I + Bg + R1 G: (4:2:80) ~ Since the problem (4.2.53) is equivalent to (4.2.67), the problem (4.2.53) is equivalent to that of nding a function w satisfying the relations ~ w 2 V2 w = : (4:2:81) Now let 0 be an eigenvalue of the problem (4.2.61) of odd multiplicity. By virtue of Lemma 4.2.5, the problem (4.2.61) has a discrete spectrum. Since only points of the spectrum of the problem (4.2.61) can be bifurcation points of the problem (4.2.53) (see Theorem 4.2.1), for any xed, su ciently small " > 0, there ~ exists a ball d" in V2 centered at w = in which the problem

w 2 d"

0 ;" w

4.2. Problem of stability of a thin plate

241

has a unique solution w = . Similarly, the problem

w 2 d"

0 +" w

has a unique solution w = . Lemmas 4.2.3, 4.2.4, and (4.2.65) imply that

F : w ! Fw = Bg w + R1 Gw ~
~ ~ is a compact, continuously Frechet di erentiable mapping of V2 into V2 , and the following equality holds F 0 = Bg : ~

L is the boundary of an arbitrary open subset of d" which contains , then the rotations of the mappings 0 ;" and 0 +" on L have di erent signs, i.e., the mappings are not homotopic on L. Let us show that there exists a pair (w ~) such ~
that by

Now, the Leray{Schauder theorem (see Krasnoselskii (1956)) implies that, if

w2L ~

~ 2( 0 ;"

0 + ")

~ ~w =

(4:2:82) ~ 0 1] ! V2 given

Let us assume the contrary. Then, the function 't (w): L

(w t) ! 't (w) = w + t( 0 + ") + (1 ; t)( 0 ; ")]Bg w + R1 Gw ~

does not vanish at any point, and '0 (w) = 0 ;" w, '1 (w) = 0 +" w. Hence, the mappings 0 ;" and 0 +" are homotopic on L. But we have established that this is not the case. Thus, (4.2.82) holds, and the theorem is proven.

Conclusion 1 It is natural to connect the loss of stability of the plate with the

least positive bifurcation point of the problem (4.2.53). Only the eigenvalues of the problem (4.2.61) can be bifurcation points of the problem (4.2.53), and every eigenvalue of the problem (4.2.61) of odd multiplicity is a bifurcation point of (4.2.53). The question whether an eigenvalue of (4.2.61) of even multiplicity is a bifurcation point of (4.2.53) demands an additional investigation. However, below we will always connect with the loss of stability of the plate the least positive eigenvalue of the problem (4.2.61), at which the form bg is given ~ by the expression (4.2.51), the latter being determined, in turn, by the solution of the problem (4.2.49). From the physical point of view, this means that the form bg is determined by inner forces, which depend on the function of exterior ~ longitudinal forces Q and on the law of change of the thickness of the plate (i.e., on the function h(x y)).

242

Chapter 4. Direct problems for plates and shells

z

2b
x

l Q1 ( y )
a

Figure 4.2.1: Rectangular plate clamped along the edge x = 0 and subject to compressing forces Q1 (y) Let a rectangular plate of constant thickness be clamped along the edge x = 0 and be subject to longitudinal compressing forces Q1 (y) acting along the edge x = a (see Fig. 4.2.1). According to (4.2.2) and (4.2.5), the stored energy of the plate is given by (!) =
2 (! ) + 1 (2) 2 a (w

4.2.6 An example of nonexistence of stable solutions

w) ;

Zb

;b

Q1(y)u(a y) dy

! = (u v w ) :
(4.2.83)
2 (! )

Here, the quadratic form a(2) (w w) is de ned by (4.2.31) and the functional is given by 1 Z a Z b c (" (!))2 + 2c " (!)" (!) (!) = 2 11 11 12 11 22 2
0
;b

+ c22 ("22 (!))2 + c33 ("12 (!))2 dx dy: (4.2.84)

The functional (4.2.83) is considered in the space V = V1 V2 (! = (u v w) 2 V , (u v) 2 V1 , w 2 V2 ). In the case under examination, the spaces V1 and V2 are the following ones

V1 = g j g = (g1 g2 ) g 2 W21 ( )

g(0 y) = 0

4.2. Problem of stability of a thin plate

243

where = (x y) j 0 < x < a ;b < y < b . The space V2 is easily veri ed to meet the condition (4.2.11). Introduce the function

@q V2 = q j q 2 W22 ( ) q(0 y) = @x (0 y) = 0 !c = ; 2 c2 x3 0 cx2 3

;

c 2 R:

(4:2:85)

Obviously, i.e.,
;

!c 2 V

c2R cx2 2 V2 c 2 R:

Due to (4.2.1) and (4.2.85), for the strain components generated by the function !c, we have "11 (!c ) = 0 "22 (!c ) = 0 "12 (!c) = 0: (4:2:86) Now, (4.2.83){(4.2.86) yield 1 c2 a(2) (x2 x2 ) + 2 a3 c2 Z b Q (y) dy: (!c ) = 2 1 3 ;b (4:2:87)

; 2 c2 x3 0 2 V1 3

Since Q1 (y) is a compressing load (see Fig. 4.2.1) and > 0, we get Q1 (y) 0. Hence, 2 a3 Z b Q (y) dy = k < 0: (4:2:88) 1 1 3 Because a(2) (w0 Assume that Using (4.2.87){(4.2.90), we get

w ) is a coercive form on V2 V2 (see Theorem 4.1.2), we conclude

00

;b

1 a(2) (x2 x2 ) = k > 0: 2 2

(4:2:89) (4:2:90) (4:2:91)

k > jk2 j :
1

(!c ) = c2 (k2 + k1 ) < 0: Therefore, if satis es the inequality (4.2.90), then

c!1

lim

(!c ) = ; 1:

Thus, if the load is su ciently large, that is, the inequality (4.2.90) holds, then the functional (4.2.83) is not bounded below in the space V , and there is no stable equilibrium position of the plate, although there exists at least one stationary

244

Chapter 4. Direct problems for plates and shells

Here, 1 is the least positive eigenvalue of the problem (4.2.61). Indeed, by (4.2.92), the bilinear form bg takes the form ~

point of the functional (4.2.83), which is ! = ( g ), where g is the solution of the ~ ~ problem (4.2.49). Suppose that Q1 (y) = ;c1 c1 = const > 0: (4:2:92) Let us show that then k2 (4:2:93) jk j 1 :
1

bg (w0 w00 ) = ;c1 ~

Z aZ b
0

@w0 @w00 dx: ;b @x @x

(4:2:94)

(This formula is actually somewhat inaccurate, because it does not take into account the in uence of the free edges of the plate. Nevertheless, for our purposes, it is precise enough if a is much less than b.) In the case under consideration, all the eigenvalues of the problem (4.2.61) are positive and the least eigenvalue 1 is given by a(2)(w w) (4:2:95) 1 = winf ; b (w w) : 2V2 g ~ Making use of (4.2.89) and (4.2.94), after a simple calculation, we obtain
(2) 2 2 ; ab (xx xx ) = jk2 j k1 g 2 2 ~

(4:2:96)

and (4.2.96) imply (4.2.93). Thus, if the inequality (4.2.90) holds, then the load exceeds the critical value, which is connected with the rst positive eigenvalue of problem (4.2.61).

k1 being de ned by formulas (4.2.88) and (4.2.92). Since w = x2 2 V2 , (4.2.95)

4.3 Model of the three-layered plate ignoring shears in the middle layer
4.3.1 Basic relations
A three-layered plate consists of two thin exterior layers, which are made of a strong material (the so-called carrier layers), and of a comparatively light, nonstrong middle layer (the so-called ller), the latter ensures the joint work of the exterior layers. Consider the three-layered plate whose middle layer is of thickness t0 (x y) and two exterior layers are of thickness h(x y) (see Fig. 4.3.1). We suppose that h is much less than t0 (h t0 ) and that the material of the middle layer is much more exible than the material of the exterior layers. In

4.3. Model of the three-layered plate ignoring shears in the middle layer

245

h
t0

y h

Figure 4.3.1: Three-layered plate this case, the shearing stresses perceive mainly the middle layer, and the bending stresses perceive mainly the exterior ones. Suppose also that, in the transversal direction, the elasticity modulus of the material of the middle layer is in nitely large. The material of the middle layer is usually light, so that the mass of the plate is concentrated in the exterior layers. This is why, in solving optimization problems for the three-layered plates, the control is usually the function h determining the thickness of the carrier layers. In what follows, we assume that the equality t0 + h = const (4:3:1) determining the parallelism of the midplanes of the carrier layers holds. The Kirchho hypotheses are supposed to be ful lled for the three-layered plate as a whole. Then, the strain components "1, "2 , "3 are expressed by the formulas (4.1.1). The layers are assumed to be made of orthotropic materials, so that the relations (4.1.8) between stresses and strains are valid, and moreover, E11 = E12 = E21 = E22 = G = 0 for the inner layer, and the elasticity characteristics of the exterior layers coincide. Taking into account that h t0 , from (4.1.3) and (4.1.8) we deduce the following relations for the bending moments M1 , M2 and for the torque M3 :

M1 =

Z t20 +h
2

; t0 ;h

z 1 (z ) dz ; t0 + h h 1 ; t0 + h + t0 + h h 2 2 2 = D1 1 + D2 2

t0 + h
2 (4.3.2)

246 where

Chapter 4. Direct problems for plates and shells

2 D1 = E11 (t02+ h) h 2 D12 = E12 (t02+ h) h

(4:3:3)

E11 , E12 are the elasticity characteristics of the exterior layers for which (4.1.9)
holds true. Similarly, we have

M2 = M3 =
Here

Z t20 +h Z t20 +h
0 ; t2 ;h 0 ; t2 ;h

z 2 (z ) dz D21 1 + D2 z 3 (z ) dz D3 3 :

(4.3.4)

2 D21 = E21 (t02+ h) h = D12 2 D = E22 (t0 + h) h

2 D3 = G(t0 + h)2 h

(4.3.5)

and E21 , E22 , G are the elasticity characteristics of the exterior layers. The strain energy of the three-layered plate is determined by the relation (4.1.12), where D1 , D2 , D3 , and D12 , are expressed by the formulas (4.3.3), (4.3.5), and Mi 's are described by the right hand sides of (4.3.2) and (4.3.4).

The bilinear form ah (u v) for the three-layered plate is determined by (4.1.24), D1 , D2 , D3 , and D12 being de ned by the formulas (4.1.9), (4.3.3), and (4.3.5). Suppose that

4.3.2 Problems of the bending and of the free exural oscillations

E1 E2 G are positive constants, i are constants, 0 i < 1 i = 1 2 2 E1 = 1 E2 h 2 Yp t0 c1 > 0 t0 + h = c2

(4.3.6) (4.3.7) (4.3.8) (4.3.9) (4.3.10) (4:3:11)

where

In the same way as in subsec. 4.1.3, one proves the following

c1 and c2 are positive numbers.

Theorem 4.3.1 Let the conditions (4.3.6){(4.3.11) hold true, and let a set Yp be determined by the formulas (2.1.2) and (2.1.3). Assume that V is a closed subspace

4.3. Model of the three-layered plate ignoring shears in the middle layer

247

of W22 ( ) satisfying (4.1.21). Then, the bilinear form ah (u v) de ned by (4.1.9), (4.1.24), (4.3.3), and (4.3.5) is symmetric, uniformly continuous and coercive in h 2 Yp on V V , i.e., ah (u v) = ah (v u) u v2V (4.3.12) jah (u v)j c3 kukV kvkV u v 2 V h 2 Yp (4.3.13) 2 ah (u u) c4 kukV u 2 V h 2 Yp (4.3.14) c3 and c4 being positive numbers. Remark 4.3.1 The condition (4.3.10) means the parallelity of the midplanes of the exterior layers (see Fig. 4.3.1). It is easy to verify that Theorem 4.3.1 remains valid if (4.3.10) is not true. The problem of the bending of the three-layered plate reduces to the solution of the equation (4.1.29) for the corresponding form ah , where f 2 V , f being a load. The space V for the three-layered plate is chosen just as for the one-layered plate (see subsec. 4.1.4). In particular, for the space V of the form (4.1.31), f can be represented in the form (4.1.36). By the Riesz theorem, or by the Lax{Milgram theorem, under the suppositions of Theorem 4.3.1, for every load f 2 V , the problem (4.1.29) corresponding to the bending of the three-layered plate has a unique solution. We proceed now consider the problem of free oscillations of the three-layered plate. A function U (x y t) determining the de ection of the three-layered plate, which depends on the coordinates x, y and time t, is given as the solution of the following problem U 2 C 2 (0 1) V ]
ZZ Z t20 +h
t0
2

@ 2 U v + Z 20 Az 2 dz + Z ; 20 Az 2 dz (t0 0 + 2h ) @t2 0 0 0 ; t2 ;h; t2

t t

+ where

Az 2 dz dx dy + ah (U v) = 0

v 2 V t 2 (0 1)

(4.3.15) (4:3:16)

Here, 0 and are the densities of the material of the interior and exterior layers. The rst term under the sign of integral in (4.3.15) determines the work of the inertia forces on virtual transversal displacements, and the following terms, according to the Kirchho hypothesis, de nes the work of the inertia forces on virtual longitudinal displacements (see subsec. 4.1.5). By integrating in (4.3.15) in z , we arrive to the following problem U 2 C 2 ((0 1) V )

@ 3 U @v @ 3 A = @t2 @x @x + @t2U @v : @y @y

248
ZZ

Chapter 4. Direct problems for plates and shells

2 1 2 (t0 0 + 2h ) @ U v + 12 0 t3 A + h2 t0 + 1 ht2 + 3 h3 0 @t2 2 0

A dx dy
(4.3.17)

+ ah (U v) = 0

v 2 V t 2 (0 1):

We will search for the solution of the problem (4.3.17) in the form (4.1.50). Substituting (4.1.50) into (4.3.17) and taking into account (4.3.16), we get the problem ( u) 2 R V where = !2 and

ah (u v) = bh (u v)

v2V

(4:3:18)

bh (u v) =

ZZ

1 @v @v (t0 0 + 2h )uv + 12 0 t3 @u @x + @u @y 0 @x @y +

h2 t
0

1 2 2 3 0 + 2 ht0 + 3 h

@u @v + @u @v @x @x @y @y
= const > 0:

dx dy:

(4.3.19)

We suppose that

= const > 0

(4:3:20)

bh (u v) be determined by the formulas (4.3.19) and (4.3.20). Then, for all h 2 Yp , the spectral problem (4.3.18) possesses a sequence of nonzero solutions fui g V
that correspond to a sequence of eigenvalues f i g such that

Theorem 4.3.2 Let the hypotheses of Theorem 4.3.1 hold and let a bilinear form
ah (ui v) = i bh (ui v)
3

0<

form bh is symmetric, continuous, and coercive on W21 ( ) W21 ( ). Now, Theorem 4.3.2 is a consequence of the compactness of the imbedding of V into W21 ( ) and of Theorem 1.5.7.

Proof. It is easy to see that, under the hypotheses of Theorem 4.3.2, the bilinear

lim = 1 i!1 i

v2V

bh(ui uj ) = ij :

4.4 Model of the three-layered plate accounting for shears in the middle layer
Let us consider another model of the three-layered plate which accounts for shears in the middle layer (the ller). We assume that, in the ller, a straight line perpendicular to its midsurface before deformation remains straight after the deformation, but it is no longer perpendicular to the midsurface because of shears. Thus, we accept the linear law of displacements in the ller. To the thin exterior layers we

4.4.1 Basic relations

4.4. Model of the three-layered plate accounting for shears in the middle layer 249

Figure 4.4.1: Three-layered plate. A normal passing through the three layers becomes a broken line after the bending apply the straight normal (Kirchho ) hypothesis. Hence, a normal passing through the three layers becomes a broken line after the deformation (see Fig. 4.4.1). This allows one to account for shear strains of the ller. Denote by u, v, and w the components of the vector of displacements in the x, y, z axes of the midplane of the upper layer (see Fig. 4.3.1). Respectively, u1 , v1 , and w1 are the components of the vector of displacements of the midplane of the lower layer. Let also w2 be the displacement in the z axis of points of the middle layer. The functions w, w1 , and w2 are supposed to depend only on x and y, not on z , and moreover, w1 w2 w: (4:4:1) Below, we will examine problems of the bending of the three-layered plate. There one can assume that

u1 = ;u

v1 = ;v:

(4:4:2)

By the straight normal hypothesis (see Figs. 4.3.1 and 4.4.1), for the displacements uu , vu , and wu of points of the upper layer in the x, y, z axes, we have vu = v ; z + t0 + h @w : (4:4:3) wu = w uu = u ; z + t0 + h @w 2 @x 2 @y Similarly, for the displacements ul , vl , and wl of points of the lower layer, we get wl = w ul = ; u ; z ; t0 + h @w vl = ; v ; z ; t0 + h @w : 2 @x 2 @y (4:4:4) The displacements um, vm , and wm of points of the middle layer are given by

wm = w

um = ; tz 2u ; h @w @x
0

vm = ; tz 2v ; h @w : (4:4:5) @y
0

250

Chapter 4. Direct problems for plates and shells

For the strain components of the layers we use the formulas ~ "1 = @ u @x

@v ~ ~ @v ~ "2 = @y "3 = @ u + @x @y ~ ~ "5 = @ v + @w "4 = @ u + @w @z @x @z @y

(4.4.6)

u and v being the components of the vector of displacements in the corresponding ~ ~ layers which are de ned by (4.4.3){(4.4.5). The components "4 and "5 are taken
into account only for the middle layer in which the shear strains are essential. Let a three-layered plate be made of orthotropic materials, and let E1 , E2 , G, 1 , and 2 be elasticity characteristics of the exterior layers (we assume that the upper and lower layers are made of the same material), and let G1 be the shear modulus of the material of middle layer. For the stresses in the exterior layers, we have
1

= E11 "1 + E12 "2

= E21 "1 + E22 "2

= G"3

(4:4:7)

E11 , E21 , E12 , and E22 being determined by the relations (4.1.9). Substituting the values of "i , i = 1 2 3, from (4.4.6) into (4.4.7), taking to notice the formulas
(4.4.3) and (4.4.4) for the displacement components in the exterior layers, and making use of the equality (4.3.1), we obtain the following relations: for the upper layer
2 @u = E11 @x ; z + t0 + h @ w 2 @x2 @u ; z + t0 + h @ 2 w 2 = E21 @x 2 @x2 @u @v t0 + h 3 = G @y + @x ; 2 z + 2 1

@2w @x@y

@v + E12 @y ; z + t0 + h 2 @v ; z + t0 + h + E22 @y 2

@ 2w @y2 @ 2w @y2
(4.4.8)

for the lower layer

@ 2w @x2 @ 2w @x2 t0 + h @u @v 3 = ;G @y + @x + 2 z ; 2 @u = E11 ; @x ; z ; t0 + h 2 @u t0 + h 2 = E21 ; @x ; z ; 2

1

@2w : @x@y

@v + E12 ; @y ; z ; t0 + h 2 @v + E22 ; @y ; z ; t0 + h 2

@2w @y2 @2w @y2

(4.4.9)

By integrating the relations (4.4.8) in the thickness of the upper layer, we

4.4. Model of the three-layered plate accounting for shears in the middle layer 251

get the following formulas for the force per unit of length

N1u = N2u = N3u =

1 (z ) dz 0 ; t2 ;h Z ; t0 2 (z ) dz 0 ; t2 ;h Z ; t0 3 (z ) dz ; t0 ;h
2 2 2

0 ; t2

= h E11 @u + E12 @v @x @y = h E21 @u + E22 @v @x @y (4.4.10)

@v = hG @u + @x : @y

The forces N1l , N2l , and N3l in the lower layer are determined by the relations (4.4.10) with u and v replaced by ;u and ;v, respectively, i.e.,

N1u = ;N1l

N2u = ;N2l

N3u = ;N3l :

(4:4:11)

For the bending moments M1u , M2u and for the torque M3u in the upper layer with respect to the midplane of this layer (see Fig. 4.3.1), by (4.1.9) and (4.4.8), we get

M1u = M2u = M3u =

Here

0 ; t2

; 0 ;h Z ;2t20

z + t0 + h 2 z + t0 + h 2 z + t0 + h 2

0 ; t2 ;h Z ; t0 2
; t0 ;h

@2w ; D @2w 12 @y 2 @x2 @2w @2w 2 (z ) dz = ;D21 @x2 ; D2 @y 2 (4.4.12) @ 2w 3 (z ) dz = ;D3 @x@y :
1 (z ) dz = ;D1

3 D1 = 12(1E1 h ) ; 1 2 3 D = E2 h 2

12(1 ;

1 2)

E 3 D12 = D21 = 12(11; 2 h ) 1 2 1 Gh3 : D3 = 6

(4.4.13)

Similar relations can be obtained for the bending moments M1l , M2l and for the torque M3l in the lower layer with respect to the midplane of this layer, i.e.,

Mil =

Z t20 +h
t0
2

z ; t0 + h 2

i (z ) dz = Miu

i = 1 2 3:
5

(4.4.14)

In the middle layer, only shearing stresses 4 and They are given by 4 = Gm "4 5 = Gm "5

are taken into account. (4:4:15)

252

Chapter 4. Direct problems for plates and shells

Gm being the shear modulus of the material of the middle layer, and "4 , "5 being determined by the relations (4.4.6). From (4.4.5), (4.4.6), and (4.4.15), we get = G @um + @w = G @w ; 1 2u ; h @w
4 m 5

= Gm

m @x t @z @x @x 0 @vm + @w = G @w ; 1 2v ; h @w m @y t @z @y @y 0

(4.4.16)

The transversal forces in the middle layer are de ned by the formulas

Q1m = Q2m =

4 (z ) dz 0 ; t2 Z t0 5 (z ) dz ; t0
2 2

Z t20

= Gm (t0 + h) @w ; 2u @x

= Gm (t0 + h) @w ; 2v : @y

(4.4.17)

The strain energy of the three-layered plate is given by 1 ZZ dx dy Z ; 20 ( " + " + " ) dz =2 1 1 2 2 3 3 0 ; t2 ;h 1 ZZ dx dy Z 20 ( " + " ) dz +2 4 4 5 5 0 ; t2
t t

1 +2

ZZ

dx dy

Z t20 +h
t0
2

( 1 "1 + 2 "2 + 3 "3 ) dz:

(4.4.18)

Here, is the domain occupied by the midplane of the plate, which is supposed to be a Lipschitz domain. Substituting into (4.4.18) the values of i from (4.4.8), (4.4.9), and (4.4.16), taking to notice (4.4.3){(4.4.6), and denoting ! = (u v w), we get the following formula for the strain energy of the three-layered plate (!) =
ZZ

1;

E1 h

1 2

(P1 !)2 + 2 2 (P1 !)(P2 !)

3 1 2)

E + 1 ; 2 h (P2 !)2 + 12(1E1 h ;

1 2 + Gh(P3 !)2 +

(P4 !)2 + 2 2 (P4 !)(P5 !)

+ 1 Gm t0 (P7 !)2 + (P8 !)2 dx dy: 2 Here, Pi !, i = 1 2 : : : 6, are the strain components of the exterior layers, and P7 !, P8 ! are the ones of the middle layer of the plate, which are given by the

12(1 ;

E2 h3

2 1 3 2 ) (P5 !) + 3 Gh (P6 !) 2

(4.4.19)

4.4. Model of the three-layered plate accounting for shears in the middle layer 253

formulas

P1 ! = @u @x 2 P4 ! = @ w @x2 P7 ! = 1 + th @w ; 2u t0 0 @x

@v P2 ! = @v P3 ! = @u + @x @y @y 2 @2w P5 ! = @ w P6 ! = @x@y @y2 P8 ! = 1 + th @w ; 2v : (4.4.20) t0 0 @y

4.4.2 Bilinear form corresponding to the three-layered plate

Introduce the notation

W = W21 ( ) W21 ( ) W22 ( ):

(4:4:21)

If ! = (u v w) 2 W , then u 2 W21 ( ), v 2 W21 ( ), and w 2 W22 ( ). The relations (4.3.6){(4.3.11) are supposed to be valid. Then, the formulas (4.4.20) de ne the operators Pi : ! ! Pi !, and

Pi 2 L(W L2 ( ))

i = 1 2 : : : 8:

(4:4:22)

The functional (!) determined by the relations (4.4.19) and (4.4.20) is assumed to be de ned in a space V which is a subspace of W and is determined according to the way of the fastening of the plate. We suppose that

V is a closed subspace of W and the relations ! 2 V , Pi ! = 0, i = 1 2 : : : 8, imply ! = 0.

The norm of an element ! = (u v w) 2 V is de ned by

(4:4:23)

k!k2 = kuk2 21( ) + kvk2 21 ( ) + kwk2 22 ( ) : V W W W

(4:4:24)

By (4.4.19), to the strain energy of the three-layered plate there corresponds

254

Chapter 4. Direct problems for plates and shells

the symmetric bilinear form on V

ah (!0 !00 )
=2

ZZ

1;

E1 h

1 2

(P1 !0 )(P1 !00 ) + 2 ((P1 !0 )(P2 !00 ) + (P2 !0 )(P1 !00 ))

E + 1 ; 2 h (P2 !0 )(P2 !00 ) + Gh(P3 !0 )(P3 !00 )

+ 12(1 ; + 12(1 ;

E1 h3

1 2

1 2) 3 E2 h 1

(P4 !0 )(P4 !00 ) + 2 ((P4 !0 )(P5 !00 ) + (P5 !0)(P4 !00 ))

+ 1 Gm t0 (P7 !0 )(P7 !00 ) + (P8 !0 )(P8 !00 ) 2

1 3 0 00 0 00 ) (P5 ! )(P5 ! ) + 3 Gh (P6 ! )(P6 ! ) 2

dx dy

!0 !00 2 V:
(4.4.25)

(4.3.6){(4.3.11) hold, let a set Yp be de ned by the formulas (2.1.2) and (2.1.3), and let Gm = const > 0: (4:4:26) Suppose that the space V meets the condition (4.4.23) and is equipped with the norm (4.4.24). Then, the bilinear form ah (!0 !00 ) determined by the relations (4.4.20) and (4.4.25), is symmetric, uniformly continuous and uniformly coercive in h 2 Yp on V V , i.e.,

Theorem 4.4.1 Let be a bounded Lipcshitz domain in R2 , let the conditions

ah (!0 !00 ) = ah(!00 !0 ) !0 !00 2 V jah (!0 !00 )j ck!0kV k!00 kV !0 !00 2 V h 2 Yp ah (! !) c1 k!k2 ! 2 V h 2 Yp V c and c1 being positive numbers.

(4.4.27) (4.4.28) (4.4.29)

Proof. The equality (4.4.27) is a consequence of (4.4.25). It is easy to deduce the inequality (4.4.28) from (4.3.6){(4.3.11), (4.4.22), and (4.4.26). Making use of the inequality a2 + b2 2jabj which is valid for all a b 2 R and taking into
consideration (4.3.6){(4.3.11) together with (4.4.26), we get

ah (! !) c2

ZZ X 8

i=1

(Pi !)2 dx dy

! 2 V h 2 Yp

(4:4:30)

c2 being a positive number.

Let us show that the system of operators Pi , i = 1 2 : : : 8, is W -coercive with respect to (L2 ( ))3 (see subsec. 1.7.1).

4.4. Model of the three-layered plate accounting for shears in the middle layer 255

According to (1.7.2), the operators Pi de ned by (4.4.20) can be represented as follows X X X ! = (u v w) ! Pi ! = gi1k Dk u + gi2k Dk v + gi3k Dk w: (4:4:31)
jk j

jk j

jkj

Here

and Dk f = f if jkj = 0,

@ k1 +k2 Dk = @xk1 @yk2

k = (k1 k2 ) jkj = k1 + k2 i = 1 2 : : : 8:
(4:4:32)

gi1k gi2k gi3k 2 L1 ( )

X

To use Theorem 1.7.1, let us consider a rectangular matrix with the elements

Pir (x y ) =

jkj=lr

k k girk 1 1 2 2

i = 1 2 ::: 8 r = 1 2 3

where lr = 1 for r = 1 2, and lr = 2 for r = 3 (see (4.4.21)), and = ( 1 2 ) 2 C 2 , C being the set of complex numbers. From (4.4.20), (4.4.31), and (4.4.32), we deduce 2 3 0 0 0 1 0 2 0 0 (Pir (x y ))T = 4 0 2 1 0 0 0 0 0 5 2 2 0 0 0 1 2 12 0 0 the index T standing for transposition. It is easy to see that, for any 2 C 2 , 6= 0, the rows of the matrix (Pir (x y ))T are linearly independent, i.e., the rank of this matrix is equal to 3. Theorem 1.7.1 implies that the system of operators Pi de ned by (4.4.20) is W -coercive with respect to (L2 ( ))3 . Further, combining (4.4.23) and Theorem 1.7.3, we obtain
ZZ X 8

i=1

(Pi !)2 dx dy c3 k!k2 V

!2V

(4:4:33)

where c3 = const > 0. Finally, (4.4.30) and (4.4.33) yield (4.4.29), concluding the proof.

4.4.3 Bending of the three-layered plate

Let f be a load acting on a plate, which is identi ed with an element of the space V and let V satisfy the condition (4.4.23). The problem of the bending of the plate reduces to determining a function ! such that ~ !2V ~ ah (~ !) = (f !) ! ! 2 V: (4:4:34) Theorem 4.4.1 and the Riesz theorem imply the following

256

Chapter 4. Direct problems for plates and shells

Theorem 4.4.2 Under the suppositions of Theorem 4.4.1, for all h 2 Yp and f 2 V , there exists a unique function ! satisfying (4.4.34). ~
To study special realizations of the space V , which are connected with the way of the fastening of the plate, we have, in view of the condition (4.4.23), to de ne the following subspace

H = ! j ! 2 W Pi ! = 0 i = 1 2 : : : 8 (4:4:35) W being determined by (4.4.21) and Pi by (4.4.20). From the physical point of view, the subspace H de nes the set of vector functions ! 2 W for which all the strain components Pi ! are equal to 0, and

consequently, so is the strain energy (see (4.4.19)). With regard to (4.4.20), the relations Pi ! = 0 de ne a linear homogeneous system of partial di erential equations. It is easy to see that the general solution of this system has the form a) if h and t0 are positive numbers,

u = c1

v = c2

2 w = t + h (c1 x + c2 y) + c3 0

(4:4:36)

c1 , c2 , and c3 being constants

b) otherwise, i.e., if the thicknesses of the layers are not constant,

u=0 c4 being a constant.

v=0

w = c4

(4:4:37)

Consider the case a). (4.4.36) implies that the space H is three-dimensional, and a basis in it is formed by the vector functions f!i = (ui vi wi g3=1 given by i

x !1 = 1 0 t 2+ h 0

y !2 = 0 1 t 2+ h 0

!3 = (0 0 1): (4:4:38)

Let ! = (u v w) 2 H and !(x0 y0 ) = 0, where (x0 y0 ) 2 . Then, (4.4.36) implies that c1 = c2 = c3 = 0, i.e., ! = 0. Thus, if the space V corresponds to a fastening such that the displacements u, v, w vanish at least at one point, then the condition (4.4.23) is satis ed. It should be noted that, for the functions u, v, the condition of the fastening at a single point cannot be considered, because the function u ! u(x0 y0 ), (x0 y0 ) 2 , is not a continuous mapping of W21 ( ) into R. Thus, for the functions u, v, the condition of the fastening at a point must be replaced by that of the fastening (vanishing) of u, v on some one- or two-dimensional manifold. Let, for example, the space V be of the form

V = ! j ! 2 W !jS1 = 0

(4:4:39)

4.4. Model of the three-layered plate accounting for shears in the middle layer 257

where

S1 S

S being the boundary of the domain .

S1

ds > 0

(4:4:40)

This space satis es the condition (4.4.23). Indeed, the space V determined by (4.4.39), is a closed subspace of W . The relations ! 2 V and Pi ! = 0 imply ! 2 H , that is, ! = c1 !1 + c2 !2 + c3 !3 : (4:4:41) Since !jS1 = 0, we get, by the above argument, that c1 = c2 = c3 = 0, i.e., ! = 0. Moreover, since the function w ! w(x0 y0 ), where (x0 y0 ) 2 , is a continuous mapping of W22 ( ) into R, the space V of the form

V = ! = (u v w) j ! 2 W ujS1 = vjS1 = 0 w(x0 y0 ) = 0

(4:4:42)

meets (4.4.23), too. Consider now the case b) when the layers of the plate are of variable thickness. (4.4.37) yields that the subspace H is one-dimensional, and the function

! = (0 0 1)

(4:4:43)

forms a basis in it. Thus, for the case b), in order that the condition (4.4.23) be satis ed, it is su cient that the space V ensures a fastening of the plate such that w = 0 at least at one point, i.e., the space V determined by

V = ! = (u v w) j ! 2 W w(x0 y0 ) = 0
where (x0 y0 ) 2 , meets (4.4.23) in the case b).

(4:4:44)

lem of the bending of the three-layered plate to exist, it is necessary that the plate is fastened, i.e., it has no rigid displacements. But the space V de ned by (4.4.44) does not ensure the rigid fastening of the plate, while for this space there exists a unique solution of the bending problem (the problem (4.4.34)). This is explained by the roughness of the equations of the three-layered plate, since these were de ned from the hypotheses connected with displacements of points of the plate. Nevertheless, this fact is not a series obstacle in solution of applied problems one can always choose the space V in such a way that it corresponds to a rigid fastening of the plate (in particular, this is the case for the space V from (4.4.39) if S1 is not an intercept of a straight line). Let a space V be de ned by (4.4.39). An element f 2 V corresponding to the load can, for example, be represented in the form
ZZ Z

Remark 4.4.1 From the physical point of view, for a unique solution of the prob-

! = (u v w)

(f !) =

(f1 u + f2 v + f3 w) dx dy +

S2

f4 w ds: (4:4:45)

258

Chapter 4. Direct problems for plates and shells

Here f1 , f2 , and f3 are the components of vector function of the surface forces, f4 is the cutting force, S2 = S n S1 , and

f1 f2 f3 2 L2 ( ) f4 2 L2 (S2 )

S2

ds > 0:

(4:4:46)

The element f determined by (4.4.45) and (4.4.46) belongs to the space V , and for this element, in virtue of Theorem 4.4.2, there exists a unique function ! ~ solving the problem (4.4.34).

4.4.4 Natural oscillations of three-layered plate

A function U (x y t) = (U1 (x y t) U2 (x y t) U3 (x y t)) determining displacements of the three-layered plate depending on the coordinates x, y and time t during free oscillations is de ned as the solution of the problem
ZZ

U 2 C 2 ((0 1) V ) 2U 2U 2U (t0 0 + 2 h) @@t23 w0 + 2 h @@t21 u0 + @@t22 v0 2U 0 @3 + t0 0 2 @@t21 ; h @t2U3 2u0 ; h @w 12 @x @x 2U 0 @3 + t0 0 2 @@t22 ; h @t2U3 2v0 ; h @w dx dy + ah (U !0) = 0 12 @y @y t 2 (0 1) !0 2 V !0 = (u0 v0 w0 ): (4.4.47)

Here, the rst term under the sign of integral corresponds to the work of the inertia forces on virtual transversal displacements, the second one corresponds to the work of the inertia forces of the exterior layers on virtual longitudinal displacements, and the third and fourth terms de ne the work of the inertia forces of the interior layer on virtual longitudinal displacements. Notice that the latter two terms could be obtained by using (4.4.5) and integrating the elementary work in z from ; t20 to t20 in determining the work of the inertia forces of the exterior layers on virtual longitudinal displacements, we neglect the rotation inertia of these layers, i.e., the second terms in (4.4.3) and (4.4.4), which is justi ed by the fact that the thickness of the exterior layers is supposed to be small. Further, in (4.4.47), and 0 are the density of the material of the exterior and interior layers, respectively (recall that we suppose that both exterior layers are made of the same material) U1 and U2 are displacements of points of the midplane of the upper layer along the x and y axes the corresponding displacements in the lower layer are ;U1 and ;U2. We search for the solution of the problem (4.4.47) in the form

U (x y t) = (c1 cos t + c2 sin t)!(x y)

!2V

(4:4:48)

4.4. Model of the three-layered plate accounting for shears in the middle layer 259

c1 and c2 being constants. !

Here

By substituting (4.4.48) into (4.4.47), we get the following equation for and

!2V

2R
ZZ

ah (! !0 ) = bh (! !0)
=
2

!0 2 V:

(4:4:49) (4.4.50)

bh(! !0 ) =

(t0 0 + 2 h)ww0 + 2 h(uu0 + vv0 ) + t0 0 2u ; h @w 12 @x + t0 0 2v ; h @w 12 @y 2u0 ; h @w @x

0

2v0 ; h @w @y

dx dy

(4.4.51) For the three-layered plates, the density of the material of the interior layer is much less than that of the exterior layers. Hence, in (4.4.51), one can often neglect the terms with the multiplier 0 . Then, the form bh takes the form

! = (u v w) !0 = (u0 v0 w0 ):

bh (! ! ) =
0

ZZ

2 h(uu0 + vv0 + ww0 ) dx dy:

(4:4:52)

bh be de ned by (4.4.51) or (4.4.52), and 0 being positive constants. Then, for all h 2 Yp , the spectral problem (4.4.49) has a sequence of nonzero solutions f!i g1 V which correspond to a sequence of eigenvalues f i g1 such that i=1 i=1 ah (!i !) = i bh (!i !) !2V (4.4.53) 0< 1 2 lim = 1: (4.4.54) i!1 i

Theorem 4.4.3 Let the hypotheses of Theorem 4.4.1 be ful lled and let a form

Each eigenvalue in the sequence f i g1 appears as many times as its multiplicity, i=1 and the multiplicity of each eigenvalue is nite. ~ Proof. Introduce the Hilbert space V coinciding as a set with the space V and with the scalar product and norm generated by the form ah . By virtue of Theo~ rem 4.4.1, being considered as a topological space, V coincides with V , and the norm generated by the form ah is equivalent to the norm of the space V de ned by (4.4.24). Obviously, the bilinear forms bh from (4.4.51) and (4.4.52) are symmetric and continuous on the space H = L2( ) L2 ( ) W21 ( ) (if ! = (u v w), ! 2 H , then u 2 L2 ( ), v 2 L2 ( ), and w 2 W21 ( )). Moreover, the forms (4.4.51) and (4.4.52) satisfy the estimate bh (! !) ck!k2L2( ))3 !2H (4:4:55) (

260

Chapter 4. Direct problems for plates and shells

theorem follows from Theorem 1.5.8.

c being a positive number. (4.4.55) imply all the eigenvalues of the problem (4.4.49) ~ to be positive. Now, in view of the compactness of the imbedding of V into H , the

4.5 Basic relations of the shell theory

In this section, we brie y set forth some general facts about shells. A detailed exposition of the shell theory can be found, e.g., in Wang (1953), Timoshenko and Woinowsky-Krieger (1959), Ambartsumian (1974), Niordson (1985). A shell is a body bounded by two surfaces the distance between which is small as compared to the other dimensions. The set of the points equidistant from the surfaces of the shell is called the midsurface of the shell. De ne on the midsurface of the shell a curvilinear orthogonal coordinate system ( ) such that the coordinate lines, i.e., the lines = const, = const coincide with the principal curvature lines of the midsurface (see Fig. 4.5.1).
g
(a b ,g , )

b = co ns

a = con st

Figure 4.5.1: Curvilinear orthogonal coordinate system on the midsurface of the shell In other words, the midsurface of the shell is a connected two-dimensional manifold with edge in R3 , and it is de ned by the parameters , , i.e., there exists a function f mapping a set R2 into R3 , ( )2 ( ) ! f( ) = : y( z(
8 < x(

)= ) : )

(4:5:1)

Here, x, y, and z are the coordinates in R3 of points of the midsurface. The parameters and are called curvilinear coordinates. Supposing that f is a bijection of onto its range f ( ), we identify a point ( ) 2 with the point f ( )

4.5. Basic relations of the shell theory

261

of the midsurface. In the chosen system of curvilinear coordinates, the midsurface is characterized by the principle curvature radii R1 ( ), R2 ( ), and by the coe cients of the rst quadratic form A( ), B ( ), by means of which the square of the length of a linear element of the midsurface is represented as follows:

ds2 = A2 d 2 + B 2 d 2 :

(4:5:2)

To determine location of points of the shell which are not on the midsurface, one introduces the third coordinate line which is normal to the midsurface and determines the distance in the normal direction between a point ( ) of the midsurface and the point ( ) (Fig. 4.5.1). The classical shell theory is based on the Kirchho hypotheses. Displacements of points of midsurface of the shell are characterized by the components u, v, w of the vector of displacements !, i.e., ! = (u v w), where u, v, w are displacements along the coordinate lines , , , respectively. (More precisely, u, v, w are the components of the expansion of the vector of displacements in three mutually orthogonal unit vectors i, j , k, where i and j are tangential vectors to the coordinate lines , , respectively, and k is the normal vector at a given point, i, j , k being directed according to the growth of coordinates.) The components of displacements of points of the shell located at a distance from the midsurface are denoted by u , v , w and determined by the components of displacements of the midsurface by the formulas

u =u+

1 @w u 1 = ;A @ + R 1

v =v+

w =w 1 @w + v : 2 = ;B @ R2
2

(4.5.3)

h , and h|the thickness of the shell|is a function of and . Here ; h 2 2 Similarly, the strain components in an arbitrary point of the shell are de ned by those of the midsurface:

"1 = "1 +

"2 = "2 +

"12 = "12 + 2

12 :

(4:5:4)

Here, "1 , "2 , "12 are the components of the tangential strain, and 1 , 2 , 12 are the components of the exural and torsional strains of the midsurface, which are given by the formulas (see, e.g., Wang (1953), Novozhilov (1951)) 1 @u 1 w "1 = A @ + AB @A v + R @ 1 1 @v + 1 @B u + w "2 = B @ AB @ R2 B @ v +A @ u "12 = A @ B B @ A

262

Chapter 4. Direct problems for plates and shells

1 @w u 1 @A 1 @w v (4.5.5) A @ ; R1 ; AB @ B @ ; R2 1 @w v 1 @B 1 @w u B @ ; R2 ; AB @ A @ ; R1 1 @ 2 w 1 @A @w 1 @B @w 12 = ; AB @ @ ; A @ @ ; B @ @ 1 1 @v 1 1 1 @u 1 + R B @ ; AB @A u + R A @ ; AB @B v : @ @ 1 2

@ 1 @ 2 = ;B @

1 = ;A

1 @

We assume that the shell is made of an orthotropic material, so that at every point of the shell all the three principal directions of elasticity coincide with the direction of the coordinate lines (such shells are referred to as orthotropic ones). In this case, the relations between strains and stresses take the form
1

= E11 "1 + E12 "2

ZZ

= E21 "1 + E22 "2

12

= G"12

(4:5:6)

Eik being de ned by (4.1.9).

1 =2

The strain energy is given by

d d

Z h 2

;h 2

( 1 "1 + 2 "2 +

12 "12 ) H1 H2 d

(4:5:7)

Here, h is the function of the thickness of the shell, and H1 , H2 are the Lame coe cients H2 = B (1 + R ): (4:5:8) H1 = A(1 + R ) By substituting into (4.5.7) the relations for the strains and stresses from (4.5.4), (4.5.6), and by integrating in , we get the following formula for the strain energy of the orthotropic shell 1 Z Z ;c "2 + 2c " " + c "2 + c "2 =2 11 1 12 1 2 22 2 66 12 + D11 2 + 2D12 1 Here
2 2 1 2 + D22 2 + 4D66 12 1 2

AB d d :(4.5.9)

cik = hEik E22 = 1 ;E2

3 Dik = h Eik 12

1 2

E12 = 2 E11 = 1 E22

E11 = 1 ;E1

E66 = G:

1 2

(4.5.10)

Notice that, when integrating in in the expressions for H1 and H2 (see (4.5.8)), we neglected the terms R1 and R2 , since they are small as compared to 1.

4.6. Shells of revolution

263

4.6 Shells of revolution

Shells whose midsurface is a surface of revolution have numerous applications in technics. Such surfaces are formed by the revolution of an arbitrary at curve around an axis lying in the same plane. This curve is referred to as a meridian. The principal curvature lines of the shell of revolution are meridians and parallels, the latter are the lines of intersection of the surface with the planes z = const (see Fig. 4.6.1). Denote by r the distance between a point of the surface and the Oz axis. Then, the equation of the meridian has the form r = r(z ), where z 2 0 L], and the location of the meridian is determined by the angle ' which is counted from the xOz plane. Denote by X the surface of revolution that has the equation of meridian r = r(z ), where 0 < z < L, i.e., X = (x y z ) j 0 < z < L x = r(z ) cos ' y = r(z ) sin ' 0 ' < 2 :
z

4.6.1 Deformations and functional spaces

r ( z)

j
x

Figure 4.6.1: Shell of revolution Let be an open rectangle in R2 determined by = (z ') j 0 < z < L 0 < ' < 2 : De ne a mapping f of the set into R3 by the formula 8 > x = r(z ) cos ' < (z ') ! f (z ') = > y = r(z ) sin ' :z =z (4:6:1)

(4:6:2)

264

Chapter 4. Direct problems for plates and shells

r(z ) being a smooth function. The mapping f is a homeomorphism of the set onto its range |an open set in X that is the complement of the intersection of X and the half-plane (x y z ) j y = 0 x 0 ;1 < z < 1 z and ' being the curvilinear coordinates on X (we stress that the coordinate z

in R3 is a usual Cartesian coordinate, while for the midsurface it is a curvilinear coordinate). The mapping (4.6.2) is a map of the set . The coe cients of the rst quadratic form and the curvature radii of the surface of revolution are given by

dr A = 1 + dz d2 R1 = ; A3 dzr 2

1 2 2
;1

B=r R2 = rA:

(4.6.3) (4.6.4)

Let u, v, w stand for the components of the vector of displacements ! along the coordinate lines z , ' and the normal to the surface , respectively (see Fig. 4.6.1). The strain components of the midsurface from (4.5.5) in the case under investigation have the form " (!) = 1 @u + w " (!) = 1 @v + u @B + w
2 A @z R1 B @' AB @z R2 1 @v + 1 @u ; v @B 1 @ 1 @w u "12 (!) = A @z B @' AB @z 1 (! ) = ; A @z A @z ; R 1 1 @ 1 @w ; v ; 1 @B 1 @w ; u 2 (! ) = ; B @' B @' R AB @z A @z R1 2 1 @ 2w ; 1 @B @w + 1 @u + 1 @v ; 1 @B v : 12 (! ) = ; AB @z@' B @z @' R1 B @' R2 A @z B @z 1

(4.6.5)

Let g be a function de ned on the strip

Q = (z ') j 0 < z < l ;1 < ' < 1

and periodical in ' with period 2 , i.e.,

(4:6:6)

g(z ') = g(z ' + 2k ) (4:6:7) for any integer k and any z 2 (0 L). f For p 1, m 2 N and given by (4.6.1), de ne Wpm ( ) to be the space that consists of functions g de ned on the strip Q, satisfying the periodicity condition (4.6.7), and whose restriction on every bounded set 1 such that Q 1

4.6. Shells of revolution

265

f belongs to Wpm ( 1 ), the Wpm ( ) norm of g is equal to the Wpm ( ) norm of the restriction of g on . Now let f f f W = W21 ( ) W21 ( ) W22 ( ): (4:6:8) De ne operators Pi 2 L (W L2 ( )) i = 1 2 ::: 6 (4:6:9) by the formulas

P1 ! = "1 (!) P4 ! = 1 (!)

P2 ! = "2 (!) P 5 ! = 2 (! )

P3 ! = "12 (!) P6 ! = 12 (!)

(4.6.10)

where ! = (u v w) and the right hand sides in (4.6.10) are determined by (4.6.5). Introduce a space V such that

V is a closed subspace of W and the conditions ! 2 V , Pi ! = 0, i = 1 2 : : : 6, imply that ! = 0.

The norm of ! = (u v w) 2 V is evidently given by the formula

(4:6:11) (4:6:12)

k!k2 = kuk2 21( ) + kvk2 21 ( ) + kwk2 22 ( ) : V W W W

To the strain energy of a shell of revolution there corresponds the following bilinear form on V V

4.6.2 The bilinear form ah

ah (!0 !00 ) =
Z LZ 2 n
0 0

c11 (P1 !0 )(P1 !00 ) + c12 (P1 !0 )(P2 !00 ) + (P2 !0 )(P1 !00 ) + c22 (P2 !0 )(P2 !00 ) + c66 (P3 !0 )(P3 !00 ) + D11 (P4 !0 )(P4 !00 ) + D12 (P4 !0 )(P5 !00 ) + (P5 !0 )(P4 !00 ) + D22 (P5 !0 )(P5 !00 ) o + 4D66(P6 !0 )(P6 !00 ) AB dz d':

(4.6.13)

Here, !0 !00 2 V , and the coe cients cik , Dik are de ned by (4.5.10). Using (4.6.10), one can easily see that

ah (! !) = 2 (!)

!2V

(4:6:14)

(!) being the strain energy corresponding to the function of displacements ! that is determined by (4.5.9).

266

Chapter 4. Direct problems for plates and shells

Theorem 4.6.1 Let the conditions (4.3.6){(4.3.8) be ful lled, and let a set Yp be de ned by (2.1.2) and (2.1.3). Assume that the space V satis es (4.6.11) and is equipped with the norm (4.6.12). Let also
d3 r 2 L (0 L) r 2 C 2 ( 0 L]) (4.6.15) dz 3 1 r (z ) c0 z 2 0 L] c0 = const > 0: (4.6.16) Then, the bilinear form ah determined by the formulas (4.5.10), (4.6.3){

(4.6.5), (4.6.10), and (4.6.13) is symmetric, uniformly continuous and uniformly coercive in h 2 Yp on V V , i.e., (4.4.27){(4.4.29) hold, c and c1 being positive constants. (4.4.28). Let us prove (4.4.29). Making use of the inequality

Proof. The relation (4.4.27) is a consequence of (4.6.13). It is easy to verify

a2 + b2 2jabj ah (! ! ) c2
Z LZ 2 X 6
0 0

a b2R ! 2 V h 2 Yp
(4:6:17)

and taking to notice (4.3.6){(4.3.8), (4.5.10), (4.6.3), (4.6.15), and (4.6.16), we get
i=1

(Pi !)2 dz d'

c2 being a positive number.

2

Let us show that the system of operators Pi , i = 1 2 : : : 6, is W -coercive with respect to (L2 ( ))3 (see subsec. 1.7.1), the space W being de ned by (4.6.8). Analogously to the proof of Theorem 4.4.1, we introduce the matrix (Pik (z '
6 0 B 2 6 6 B ;1 2 A 1 )) = 6 (AR1 );1 1 6 0 6 4 0 (BR2 );1 2 (R1 B );1 2 (AR2 );1 1

A;1

;1 ;1

Let us prove that, for all = ( 1 where = 0 L] 0 2 ], the rank of the matrix (Pik (z ' )) equals 3. Indeed, let = ( 1 2 ) 2 C 2 and 6= 0, so that at least one of the two components of the vector does not vanish. Suppose that 1 6= 0. Then, the determinant of the matrix of the third order obtained from (Pik ) by cancelling the second, fth, and sixth rows is equal to

; (AB );1 1 2 6 2 ) 2 C 2 , = 0, and for all (z ') 2 ,

2 ;A;2 1 ;2 2 ;B 2

0 0 0

7 7 7 7: 7 7 5

(4:6:18)

A;1 1 B ;1 2 (AR1 );1

A;1
0

;A

0 0

;2

2 1

4 = ;A;4 1 6= 0:

(4:6:19)

4.6. Shells of revolution

267

Now, let 2 6= 0. In the same way, we can see that the determinant of the matrix obtained from (Pik ) by cancelling the rst, fourth, and sixth rows does not vanish. Hence, the rank of the matrix (Pik (z ' )) equals 3 for all (z ') 2 and 2 C 2 , 6= 0. Now< Theorem 1.7.1 implies the system of operators Pi to be W -coercive with respect to (L2 ( ))3 , the space W being de ned by (4.6.8). Further, due to (4.6.11), using Theorem 1.7.3, we get
Z LZ 2 X 6
0 0

i=1

(Pi !)2 dz d' c3 k!k2 V

!2V

(4:6:20)

where c3 = const > 0 and k!kV is de ned by (4.6.12) Finally, (4.6.17) and (4.6.20) yield

ah (! !) ck!k2 V
where c = const > 0, concluding the proof.

! 2 V h 2 Yp

4.6.3 The subspace of functions with zero-point strain energy

In solving particular problems, the space V is determined by the way of fastening of the shell. There one faces the problem whether the condition (4.6.11) is satis ed, which in turn is connected with the problem of nding the subspace

H = ! j ! 2 W Pi ! = 0 i = 1 2 : : : 6 :

(4:6:21)

This relation shows that, for the functions from H , all the strain components vanish, and so the strain energy equals zero. This is why H is called the space of functions with zero-point strain energy. The relations Pi ! = 0, i = 1 2 : : : 6, together with (4.6.5) and (4.6.10), de ne a linear homogeneous system of partial di erential equations. The general solution of this system can be found by separation of variables, see Grigorenko (1973), Litvinov and Medvedev (1979). The solution imply H to be a six-dimensional space, and a basis in it is formed by the functions !i = (ui vi wi )

268 given by

Chapter 4. Direct problems for plates and shells

dr dr !1 = A;1 r ; dz z cos ' z sin ' ;A;1 r dz + z cos ' dr !2 = A;1 dz cos ' ; sin ' A;1 cos ' dr dr !3 = A;1 r ; dz z sin ' ;z cos ' ;A;1 r dz + z sin ' (4.6.22) dr sin ' cos ' A;1 sin ' !4 = A;1 dz dr !5 = A;1 0 ;A;1 dz !6 = (0 r 0): Theorem 4.6.2 Let ! 2 H and let !(a '1 ) = !(a '2) = !(a '3) = 0, where a 2 0 L] and 0 '1 < '2 < '3 < 2 . Then, ! = 0.
For the proof, see Litvinov and Medvedev (1979). Remark 4.6.1 Theorem 4.6.2 implies that, if a closed subspace V of the space W is de ned in such a way that the shell is fastened at least at three distinct points on some parallel, i.e., every vector function ! from V vanishes in this points, then the condition (4.6.11) is satis ed.

4.7 Shallow shells

A shell is said to be shallow if the rise of its midsurface does not exceed one fth of the least linear size of its plan. For example, the shell on Fig. 4.7.1 is shallow if 1 min(a b). For a shallow shell, according to (4.5.1), one usually sets = x, 5 = y, and de nes the midsurface by the relation z = f (x y): (4:7:1) The coe cients of the rst quadratic form A and B are set equal to one. In the strain components 1 , 2 , 12 , the terms containing u, v and their derivatives are neglected. Then, the relations (4.5.5) take the form

Here

@u w "1(!) = @x + R 1 @2w 1 (! ) = ; @x2

w "2 (!) = @v + R @y 2 @ 2w 2 (! ) = ; @y 2

@v "12 (!) = @u + @x @y @2w 12 (! ) = ; @x@y :

(4.7.2) (4:7:3)

1 = ; @2f R1 @x2

1 = ; @2f : R2 @y2

4.7. Shallow shells

z

269

y
d

Figure 4.7.1: Shallow shell

; ; We suppose that R1 1 R2 1 2 L1 ( ). However, for shallow shells, one usually sets 1 = const 1 = const: (4:7:4)

Let R2 be a bounded domain occupied by the projection of the midsurface of the shell onto the x y plane. Further, let W = W21 ( ) W21 ( ) W22 ( ): (4:7:5) De ne operators Pi 2 L(W L2 ( )) by P1 ! = "1 (!) P2 ! = "2 (!) P3 ! = "12 (!) P4 ! = 1 (!) P5 ! = 2 (!) P6 ! = 12 (!) (4.7.6) where ! = (u v w) 2 W and the right hand sides in (4.7.6) are determined by the formulas (4.7.2). Suppose also that V is a closed subspace of W such that the conditions ! 2 V , (4:7:7) Pi ! = 0, i = 1 2 : : : 6, imply ! = 0. To the strain energy of the shallow shell there corresponds the following bilinear form on V V

R1

R2

ah (!0 !00 ) =

ZZ n

c11 (P1 !0 )(P1 !00 ) + c12 (P1 !0 )(P2 !00 ) + (P2 !0 )(P1 !00 )

+ c22 (P2 !0 )(P2 !00 ) + c66 (P3 !0 )(P3 !00 ) + D11 (P4 !0 )(P4 !00 ) + D12 (P4 !0 )(P5 !00 ) + (P5 !0 )(P4 !00 ) + D22 (P5 !0 )(P5 !00 ) o + 4D66 (P6 !0 )(P6 !00 ) dx dy (4.7.8)

270

Chapter 4. Direct problems for plates and shells

cik and Dik being de ned by formulas (4.5.10). ah (! !) = 2 (!)

Taking (4.7.6) into account, it is easy to see that

!2V

where (!) is the strain energy corresponding to the function of displacements ! which is de ned by (4.5.9) with A = B = 1, = x, = y and (4.7.2).

Theorem 4.7.1 Let the conditions (4.3.6){(4.3.8) be ful lled, and let a set Yp be de ned by (2.1.2) and (2.1.3). Suppose that the space V meets (4.7.7) and is equipped with the norm (4.6.12). Let also
; R1 1 2 L1 ( ) ; R2 1 2 L1 ( ):

(4:7:9)

Then, the bilinear form ah de ned by (4.5.10), (4.7.2), (4.7.6), and (4.7.8) is symmetric, uniformly continuous and uniformly coercive in h 2 Yp on V V , i.e., the formulas (4.4.27){(4.4.29) hold true, c and c1 being positive numbers.

Theorem 4.7.1 is proven just as Theorem 4.6.1, by using Theorems 1.7.1 and 1.7.3. So, we omit the proof. We note only that the matrix (Pik ) in this case has the form. 2 3 0 0 0 1 0 2 0 0 5: (Pik (x y ))T = 4 0 2 1 0 2 ; 2 ;1 2 0 0 0 ;1 2 It is easy to see that, for all 2 C 2 , 6= 0, the rows of the matrix Pik (x y ) T are linearly independent, i.e., the rank of the matrix equals 3. Finally, we notice that, for the shallow shell, under the condition (4.7.4), the space H given by (4.6.21), i.e., the space of functions with zero-point strain energy is ve-dimensional and a basis in it is formed by the functions
2 2 ; ; ; ; !1 = R1 1x R2 1 y ;1 !2 = 2x ; 2y R2 1 xy ;x R1 R2 2 2 ; !3 = R1 1xy 2y ; 2x ;y !4 = (1 0 0) !5 = (0 1 0): R2 R1

(4.7.10)

4.8 Problems of statics of shells

We proceed study problems of stress-strain state of a shell that is subject to an exterior load. The shell is assumed to be fastened in some way, which determines the space V . The exterior load acting on the shell is identi ed with an element

4.8. Problems of statics of shells

271

f 2 V . Then, the problem of strain-state state of the shell reduces to nding a function ! = (~ v w) such that ~ u~ ~ !2V ~ ah (~ !) = (f !) ! ! 2 V:
(4:8:1) By using the Riesz theorem, we conclude that, if the hypotheses of Theorem 4.6.1 are ful lled, then the problem (4.8.1) for a shell of revolution has a unique solution for all h 2 Yp . Similarly, for a shallow shell, under the assumptions of Theorem 4.7.1, the problem (4.8.1) has a unique solution for all h 2 Yp . As an example, let us consider a shell of revolution. Let the shell be clamped at one edge. Then, the function of displacements ! = (u v w) satis es the boundary conditions

u(0 ') = 0 v(0 ') = 0 w(0 ') = 0 @w (0 ') = 0 ' 2 (0 2 ) @z

and the space V has the form
n o V = ! = (u v w) ! 2 W !(0 ') = 0 @w (0 ') = 0 ' 2 (0 2 ) (4.8.2) @z W de ned by (4.6.8). By the imbedding theorem, or by the theorem on the trace space, V is a closed subspace of W . Now, by using Theorem 4.6.2, we deduce the space V from

(4.8.2) to meet the condition (4.6.11). Further, let the load f acting on the shell be determined by the formula (f !) = +
Z2
0

Z LZ 2
0 0

(f1 u + f2 v + f3 w) AB dz d'

f4 u(L ') + f5 v(L ') + f6 w(L ') + f7 @w + f8 @w B d' (4.8.3) A @z B @'

A and B de ned by (4.6.3). Here, f1 , f2, and f3 are the components of the force acting on surface of the shell, f4 , f5 , and f6 are the components of the force acting on the edge of the shell, f7 and f8 are the bending moment and torque. We suppose that fi 2 L2((0 L) (0 2 )) fi 2 L2(0 2 ) i=1 2 3 i = 4 5 : : : 8:
(4.8.4)

Then, by the imbedding theorem and the theorem on the trace space, the element f de ned by (4.8.3) belongs to V , so that under the implied conditions the problem (4.8.1) has a unique solution.

272

Chapter 4. Direct problems for plates and shells

4.9 Free oscillations of a shell

Let ah be a bilinear form on V V generated by the strain energy of the shell let and be curvilinear coordinates, i.e., the parameters determining the midsurface of the shell. Denote by U the vector function of displacements of the midsurface during free oscillations. These displacements depend on the coordinates , and time t, i.e., U ( t) = (U1 ( t) U2 ( t) U3 ( t)) : Here, Ui ( t), i = 1 2 3, are the components of the expansion of the vector of displacements in the three mutually orthogonal unit vectors i, j , and k, where i and j are tangent to the coordinate lines and , respectively, and k is normal to the surface at a given point (see Fig. 4.5.1). The function U is a solution of the following problem

U 2 C 2 ((0 1) V ) 2U 2U 2U h @@t21 u0 + @@t22 v0 + @@t23 w0 h3 1 @ 3 1 2U 1 0 u0 + 12 A @t2U3 ; R @@t21 A @w ; R @ @ 1 1 h3 1 @ 3 U3 ; 1 @ 2 U2 1 @w0 ; v0 + 12 B @t2 @ R2 @t2 B @ R2 0 AB d d + ah(U ! ) = 0 t 2 (0 1) !0 2 V !0 = (u0 v0 w0 ):
ZZ

(4.9.1)

(4.9.2)

Here, is the density of the material of the shell and the expression under the sign of integral determines the work of the inertia forces on virtual displacements. This expression is obtained from (4.5.3) just as it was done in subsecs. 4.1.5 and 4.3.2. From the physical point of view, V is the space of virtual displacements. We search for the solution of the problem (4.9.1), (4.9.2) in the form

U ( t) = (c1 cos gt + c2 sin gt)!( ) ! = (u v w) 2 V (4.9.3) c1 and c2 being constants. By substituting (4.9.3) into (4.9.2), we get the following
eigenvalue problem

!2V
where
ZZ

2R

ah(! !0 ) = bh (! !0 )
= g2

!0 2 V

(4:9:4)

bh (! !0) =

h3 1 u h (uu0 + vv0 + ww0 ) + 12 A @w ; R @ 1

(4.9.5) 1 @w0 ; u0 A @ R1

4.10. Problem of shell stability

273

h3 1 v 1 @w0 ; v0 AB d d : + 12 B @w ; R (4.9.6) @ B @ R2 2 In particular, if the shell is thin, then in (4.9.6) we can neglect the terms containing h3 . Then, the form bh takes the form bh(! !0 ) = V
ZZ

h (uu0 + vv0 + ww0 ) AB d d :

(4:9:7)

The proof is analogous to the one of Theorem 4.4.3, and so we omit it. We notice that Theorems 4.6.1 and 4.7.1 imply the hypotheses of Theorem 4.9.1 to be ful lled, in particular, for shallow shells and shells of revolution.

Theorem 4.9.1 Let ah be a bilinear, symmetric, continuous, coercive form on V , where V is a closed subspace of W21 ( ) W21 ( ) W22 ( ) and h 2 Yp . Assume that the form bh is determined either by (4.9.6) or by (4.9.7), = const > ; ; 0, and A B A;1 B ;1 R1 1 R2 1 2 L1 ( ). Then, the spectral problem (4.9.4) has a sequence of nonzero solutions f!i g1 V which correspond to a sequence i=1 of eigenvalues f i g1 such that i=1 ah (!i !) = i bh (!i !) !2V (4.9.8) 0< 1 2 3 lim = 1: i!1 i

4.10 Problem of shell stability

4.10.1 On some approaches to stability problems

Basic relations of shell stability are obtained on the basis of nonlinear theory. The strain energy of a shell is determined by (4.5.9), except that additional nonlinear terms are inserted in the strain components "1 , "2 , "12 (see Koiter (1966), Vanin et al. (1978), Grigoliuk and Kabanov (1978)). Then, the strain energy (!) is a functional of fourth power of ! 2 V , where ! is a vector function of displacements. An exterior load acting on the shell is supposed to change proportionally to a parameter 0. Then, the stored energy of the shell is (!) = (!) ; (f !) !2V (4:10:1) where f 2 V , f being the load. When is su ciently small, there usually exists a unique stationary point ! 2 V of the functional (4.10.1), at which this functional reaches its minimum on V . The stationary point ! is the solution of the problem of the stress-strain state of the shell under the load f more exactly, ! is the function of displacements of the shell under the load f . For small , the second variation of the functional (4.10.1) is positive de nite at the point ! , i.e.,

d2 dt2

(! + th)jt=0 > 0

h 2 V h 6= 0:

(4:10:2)

274

Chapter 4. Direct problems for plates and shells

Suppose that there exists 0 such that at 0 < < 0 the condition (4.10.2) is ful lled, but for > 0 this is not the case. Then, the shell is considered to loose stability under the load 0 f . In view of this, the results on the branching of solutions of the operator equation P !=0 (4:10:3) are applied. Here, P is the Frechet derivative of the functional (4.10.1), and 0 is connected with the branch point of the equation (4.10.3) (see, for example, Srubshchik (1981)).

where "1 , "2 , and "12 are de ned in (4.5.5), and the components of the bending and torsion strains, 1 , 2 , 12 , are determined by the formulas (4.5.5). The strain energy of the shell (!) is given by (4.5.9) with "1 , "2 , "12 replaced by "1 , "2 , "12 , ^ ^ ^ respectively. Let ! be a solution of the problem (4.10.3), i.e., the stationary point of the functional (4.10.1), and let g! (!0 !00 ) be the bilinear symmetric form generated by the second variation of the functional (4.10.1) at point ! , i.e.,

To solve the stability problem, we will use an approximate approach in which the stability problem is reduced to the eigenvalue problem. When de ections of a shell are small as compared to its thickness, the components of the tangential strains of the midsurface have the form 2 2 1 1 1 "1 = "1 + 1 A @w ^ "2 = "2 + 2 B @w ^ 2 @ @ 1 @w @w "12 = "12 + AB @ @ ^ (4.10.4)

4.10.2 Reducing of the stability problem to the eigenvalue problem

(4:10:6) then the second variation of the functional (4.10.1) is not positive, i.e., the condition (4.10.2) is not satis ed. For xed !0 !00 2 V , the mapping ! ! g! (!0 !00 ) is a quadratic functional in V . The displacements and strains in the shell up to the loss of stability are supposed to be small. Then, in the formula for g! (!0 !00 ), where ! = (u v w ), one can neglect the terms containing 1 @w 2 1 @w @w : 1 @w 2 (4:10:7)

d2 (! + t!) = g (! !) ! t=0 dt2 If, for some > 0, there exists ! such that ~ !2V ~ ! 6= 0 ~ g! (~ !) = 0 !

! 2 V: !2V

(4:10:5)

A @

B @

AB @

4.10. Problem of shell stability

275

The form g! (!0 !00 ) in which the terms containing the expressions from (4.10.7) are canceled will be denoted by g! (!0 !00 ). Notice that ! ! g! (!0 !00 ) ~ ~ is an a ne mapping of V into R, !0 and !00 being xed. Taking into account the above hypothesis about displacements and strains, we de ne ! not as a stationary point of the functional of fourth power (4.10.1), but as a stationary point of the corresponding quadratic functional thus

!2V ^

! = ! ^ ah (^ !) = (f !) !

!2V

(4.10.8) (4.10.9)

where ah is the bilinear form generated by the quadratic strain energy and determined by the formulas (4.6.13) and (4.7.8) for a shell of revolution and for a shallow shell, respectively. Now, instead of (4.10.6), we get the following problem: Find 2 R for which there exists a function ! such that ~ ! 2 V ! 6= 0 ~ ~ (4.10.10) g! (~ !) = 0 ~ ! ! 2 V: By (4.10.8), we get the following representation of the form g! : ~

g! (!0 !00 ) = ah (!0 !00 ) ; bh ! (!0 !00 ) ~ (4:10:11) ^ and for xed !0 !00 2 V the mapping ! ! bh ! (!0 !00 ) is a linear functional in V .
Finally, the problem of shell stability reduces to determining the least positive eigenvalue of the following problem

2R ! 2 V ! 6= 0 ~ ~ ah (~ !) = bh ! (~ !) ! ! !2V ^
the function ! is the solution of the problem (4.10.9). ^

(4.10.12)

Under the assumption that the bilinear form ah is continuous, symmetric, and coercive on V V , we denote by V1 the Hilbert space that coincides as a set with V and the scalar product in which is generated by the form ah , V1 and V coinciding as topological spaces. Let Bh ! 2 L(V1 V1 ) be the operator generated by the bilinear form bh ! and ^ ^ de ned by the relation

4.10.3 Spectral problem (4.10.12)

ah (Bh ! ! !0 ) = bh ! (! !0 ) ^ ^

!0 2 V:

(4:10:13) (4:10:14)

Then, the problem (4.10.12) is equivalent to the following one

2R

! 2 V1 ! 6= 0 ~ ~

! ; Bh ! ! = 0: ~ ^~

276

Chapter 4. Direct problems for plates and shells

Under additional suppositions, one can show Bh ! to be a compact mapping ^ of V1 into itself for the shells of revolution and shallow shells. Hence, by applying Theorem 1.5.5, we deduce the problem (4.10.14), or, equivalently, (4.10.12), has a countable set of eigenvalues f i g1 . If 1 is the least positive eigenvalue, then i=1 1 f is the critical load under which the shell loses stability. If all the eigenvalues of the problem (4.10.12) are negative, the shell does not lose stability under all the loads f with 0 < < 1. The bilinear form bh ! (!0 !00 ) is de ned by a rather cumbersome formula, ^ which is not cited here. However, if the subcritical deformed state is momentless, then the form bh ! has the essentially simpler form ^ ZZ T1 (h !) @w0 @w00 + T2 (h !) @w0 @w00 ^ ^ 0 ! 00 ) = ; b (!
h! ^

A2 @ @ B2 @ @ 0 00 0 00 h^ + S (AB!) @w @w + @w @w AB d d @ @ @ @ !0 = (u0 v0 w0 ) !00 = (u00 v00 w00 ): (4.10.15) Here, T1 (h !), T2 (h !), and S (h !) are the forces acting in the midsurface of the ^ ^ ^
shell:

the relations (4.5.10) and (4.5.5). Notice that, if the form bh ! can be represented in the form (4.10.15), then, ^ within the limit of accuracy accepted, the forces T1 , T2, S can be determined not from the equations (4.10.9) and (4.10.16), but from the membrane (momentless) theory of shells. Example. Consider the problem of the stability of a shell of revolution. In this case, = z , = ', the coe cients A and B are de ned by the formulas (4.6.3), and is determined by (4.6.1). Let f 2 V be an exterior load acting on the shell. If the suppositions of Theorem 4.6.1 are ful lled, there exists a unique function ! ^ satisfying (4.10.9). Thus, by using the formulas (4.10.15) and (4.10.16), we de ne the bilinear form bh ! (!0 !00 ). ^ By using the compactness of the imbedding of W22 ( ) into W41 ( ), one easily sees that, in the case under investigation, the operator Bh ! de ned by (4.10.13) ^ and (4.10.15) is a compact mapping of V1 into itself (V1 being the Hilbert space containing the elements of the space V and endowed with the scalar product generated by the form ah ). Hence, there exists a countable set of eigenvalues of the problem (4.10.14) or, equivalently, of the problem (4.10.12). For a shallow shell, we have A = B = 1, = x, and = y. Similarly to the above argument, by using Theorem 4.7.1, one can conclude that, for the shallow shell, there exists a countable set of eigenvalues of the problem (4.10.12).

T1(h !) = c11 "1 (^ ) + c12 "2 (^ ) ^ ! ! T2 (h !) = c12 "1 (^ ) + c22 "2 (^ ) ^ ! ! S (h !) = c66 "12 (^ ) ^ ! (4.10.16) and the coe cients c11 , c12 , c22 , c66 and the components "1 , "2 , "12 are de ned by

4.11. Finite shear model of a shell

277

4.11 Finite shear model of a shell

4.11.1 Strain energy of an elastic shell
So far we have considered models of plates and shells based on the Kirchho hypotheses. However, a great number of composite materials have a low shear sti ness. This is why shell and plate theories which account for characteristic features of these materials, the principal one being a low shear sti ness, have been developed (see, e.g., Pelekh (1973), Vanin et al. (1978), Christensen (1979), Guz et al. (1980)). One of the most frequently used models of that kind is the nite shear model of shell, or the Timoshenko model, which is examined below. By the Timoshenko hypothesis, an element of the shell that is normal to the midsurface before deformation is no longer normal after the deformation, but rotates for some angle, remaining undistorted and of the same size. The components of displacements of points of the shell u , v , w along the coordinate lines , , (see Fig. 4.5.1) are determined through the components of displacements of the midsurface u, v, w by the formulas u =u+ 1 v =v+ 2 w =w (4:11:1) where 1 and 2 are the angles of the rotation of the normal to the midsurface, which are supposed to be independent of the components u, v, w (compare with the formulas (4.5.3)). Let ! = (u v w 1 2 ) be a vector function of displacements of the midsurface and the angles of the rotation of normals to it. The coordinate lines are supposed to coincide with the principal curvature lines of the midsurface. The strain components of the midsurface are expressed by the formulas (see, e.g., Guz et al. (1980)) " (!) = 1 @u + 1 @A v + w

where A and B are the coe cients of the rst quadratic form (see (4.5.2)), R1 and R2 are the principal curvature radii of the midsurface of the shell.

A @ AB @ R1 1 @v 1 w "2 (!) = B @ + AB @B u + R @ 2 B @ v +A @ u "12 (!) = A @ B B @ A 1 u 1 v "13 (!) = 1 + A @w ; R "23 (!) = 2 + B @w ; R @ @ 1 2 1 @ 1 + 2 @A 1 @ 2 + 1 @B 1 (! ) = A @ 2 (! ) = B @ AB @ AB @ B @ 2 +A @ 1 2 12 (!) = A @ B B@ A

(4.11.2)

278

Chapter 4. Direct problems for plates and shells

Similarly to the above reasonings (see Section 4.5), by expressing the strain components at an arbitrary point of the shell through the strain components of the midsurface (see (4.5.4)), and by integrating in the thickness of the shell, we get the following relation for the strain energy of an orthotropic shell: 1 ZZ c (" (!))2 + 2c " (!)" (!) + c (" (!))2 (!) = 2 11 1 12 1 2 22 2 + c33 ("12 (!))2 + c44 ("13 (!))2 + c55 ("23 (!))2 + D11 ( 1 (!))2 + 2D12 1 (!) 2 (!) + D22 ( 2 (!))2 + 4D33( Here,
3 cik = hEik Dik = h Eik E11 = 1 ;E1 E22 = 1 ;E2 12 1 2 1 2 E12 = 2 E11 = 1 E22 E33 = G12 E44 = G13 E55 = G23 12 (! )) 2

AB d d :
(4.11.3)

(4.11.4)

where h = h( ) is the function of the thickness of the shell, E1 , E2 , 1 , 2 , G12 , G13 , and G23 are elasticity constants of the material of the shell. The functional (4.11.3) is considered in some closed subspace V of the space ; 1 W2 ( ) 5 , V being determined by the way of the fastening of the shell. Analogously to the above, we de ne on V V a bilinear symmetric form ah (!0 !00 ). For speci c kinds of shells, under natural assumptions, one easily derives the continuity of the form ah (!0 !00 ), and, by applying Theorems 1.7.1 and 1.7.3, proves the coercivity of this form in the space V .

Remark54.11.1 A valuable circumstance is that for the shear model V ; 1 W2 ( ) , while in the classical shell theory, which uses the Kirchho hypotheses, the component w of the vector function ! = (u v w) belongs to W22 ( ). Since approximation on triangle nite elements in the space W21 ( ) is easily realized, while in the space W22 ( ) it leads to considerable obstacles (see, e.g., Ciarlet (1978)),

this circumstance allows one to use e ectively triangle nite elements for the shear model and to construct solutions of corresponding problems in the case when is an arbitrary bounded polygon in R2 . Below we present some results for the shear model of a shallow shell.

4.11. Finite shear model of a shell

279

4.11.2 Shallow shell

The bilinear form ah and the problem of statics

For a shallow shell, one usually sets = x, = y, and the midsurface of the shell is de ned by a relation z = f (x y). The coe cients of the rst quadratic form are set equal to 1. Then, the strain components take the form

w "1 (!) = @u + R @x

Here, R1 and R2 are the curvature radii of the midsurface determined by the relations (4.7.3). Let be a bounded domain occupied; by the projection of the midsurface of the shell on the (x y) plane, and let W = W21 ( ) 5 . Supposing that ; ; R1 1 R2 1 2 L1 ( ) (4:11:6) de ne operators Pi 2 L(W L2 ( ))) by the relations

@v w @v "2 (!) = @y + R "12 (!) = @u + @x @y 1 2 @w ; v @w ; u "23 (!) = 2 + @y R (4.11.5) "13 (!) = 1 + @x R 1 2 @1 @2 2 12 (!) = @ 2 + @ 1 : 1 (! ) = @x 2 (! ) = @y @x @y

P1 ! = "1 (!) P5 ! = "23 (!)

P2 ! = "2 (!) P6 ! = 1 (!)

P3 ! = "12 (!) P7 ! = 2 (!)

;

P4 ! = "13 (!) P8 ! = 12 (!):

(4.11.7) (4:11:8)

suppose also that

V1 is a closed subspace of W = W21 ( ) 5 such that the conditions ! 2 V1 , Pi ! = 0, i = 1 2 : : : 8, imply ! = 0.

;

The space V1 is equipped with the norm of the space W21 ( ) 5 . With the strain energy of the shell we connect the following bilinear form ah de ned on V1 V1 :

ah (!0 !00 ) =

ZZ n

c11 (P1 !0 )(P1 !00 ) + c12 (P1 !0 )(P2 !00 )

+ (P2 !0 )(P1 !00 ) + c22 (P2 !0 )(P2 !00 ) + c33 (P3 !0 )(P3 !00 ) + c44 (P4 !0 )(P4 !00 ) + c55 (P5 !0 )(P5 !00 ) + D11 (P6 !0 )(P6 !00 ) + D12 (P6 !0 )(P7 !00 ) + (P7 !0 )(P6 !00 ) + D22 (P7 !0 )(P7 !00 )

280

Chapter 4. Direct problems for plates and shells

+ 4D33(P8 !0 )(P8 !00 ) dx dy shell.

!0 !00 2 V1 :

(4.11.9)

Obviously, ah(! !) = 2 (!) for all ! 2 V1 , (!) is the strain energy of the

Theorem 4.11.1 Let a space V1 meet the condition (4.11.8), and let a bilinear
V1 V1 :

form ah be determined by the relations (4.11.4){(4.11.7) and (4.11.9), where E1 , E2 , G12 , G13 , and G23 are positive numbers, 1 and 2 are constants, 0 i < 1, i = 1 2, 1 E2 = 2 E1 . Then, for all h 2 Yp , Yp being de ned by the formulas (2.1.2) and (2.1.3), the bilinear form ah is symmetric, continuous, and coercive on

Proof. The symmetry and continuity of the form are obvious, and its coercivity
is proven as before, with the help of Theorems 1.7.1 and 1.7.3. The problem of stress-strain state of the shallow shell reduces to nding a function ! = (~ v w ~1 ~2 ) such that ~ u~ ~

! 2 V1 ~

ah(~ !) = (q !) !

! 2 V1

(4:11:10)

By the Riesz theorem, under the suppositions of Theorem 4.11.1, the problem (4.11.10) has a unique solution.

q being a given element of the space V1 , which corresponds to the load acting on the shell. The space V1 is de ned by the way of the fastening of the shell.

Denote by U = (U1 : : : U5 ) the vector function of displacements of the midsurface and the angles of rotation of normals during free oscillations, U be a function of coordinates x, y and time t. The function U is a solution of the following problem

Free oscillations of a shell

U 2 C 2 ((0 1) V1 ) ZZ 2U 2U 2U h @@t21 u0 + @@t22 v0 + @@t23 w0 h3 2 U 0 2 U 0 + 12 @@t24 1 + @@t25 2 dx dy + ah (U !0) = 0 0 0 t 2 (0 1) !0 = (u0 v0 w0 1 2 ) 2 V1

where is the density of the material of the shell, = const > 0. Analogously to the above, we search for the solution of the problem of free oscillations of the shell in the form

U (x y t) = (c1 cos gt + c2 sin gt) !

! 2 V1

4.11. Finite shear model of a shell

281

c1 and c2 being constants. As a result, we get the following eigenvalue problem ! 2 V1 ! 6= 0 2 R0 ah (! !0 ) = bh (! !0 ) ! 2 V1 (4.11.11)
where = g2

bh (! ! ) =
0

ZZ

h3 h(uu0 + vv0 + ww0 ) + 12 (

1 1 + 2 2)
0 0

dx dy:
(4.11.12)

Obviously, for all h 2 Yp , the bilinear form bh is continuous, symmetric, and coercive in (L2 ( ))5 . Since the imbedding of W21 ( ) into L2 ( ) is compact, under the assumptions of Theorem 4.11.1, the spectral problem (4.11.11) has a sequence of nonzero solutions f!i g1 V1 corresponding to a sequence of eigenvalues i=1 f i g1 such that i=1

ah (!i !) = i bh (!i !) ! 2 V1 0< 1 2 lim i = 1: i!1

4.11.3 A relation between the Kirchho and Timoshenko models of shell

1 = ; @x

By comparing the relations (4.7.2) and (4.11.5), one sees that the Kirchho model is obtained from the Timoshenko model when

@w

u v and in the expressions of "13 (!) and "23 (!) the terms R1 and R2 are neglected. Thus, we can consider that the bilinear form (4.11.9) is deduced from (4.7.8) when, wishing to reduce the order of derivatives in the form (4.7.8), one introduces new functions 1 and 2 connected with w by (4.11.13), and solves the problem (4.8.1) by the penalty function method. In this case, G13 and G23 are considered as parameters of the penalty. We will show the the solutions of the Tomoshenko model converge to the solution of the Kirchho model as G13 and G23 tend to in nity. Let, for example, the shell be clamped on a part S1 of the boundary S . In this case, the spaces V and V1 for the Kirchho and Tomoshenko models are de ned by

= ; @w @y

(4.11.13)

V = ! j ! = (u v w) 2 W21 ( ) W21 ( ) W22 ( ) u S1 = v S1 = w S1 = 0 @w = 0 @ S1

(4.11.14)

282

Chapter 4. Direct problems for plates and shells

= 0 on S1 (4.11.15) where 1 and 2 are the components of the unit outward normal to S1 . We de ne the norm in the space V1 as
2 2

V1 = g j g = (u v w 1 2 ) 2 W21 ( ) 5 u S1 = v S1 = w S1 = 0 1 1 +
;

kgkV1 = kuk2 21( ) + kvk2 21 ( ) + kwk2 21 ( ) + k 1 k2 21 ( ) + k 2 k2 21 ( W W W W W

1=2

(4.11.16)

Suppose that the shear modula G13 and G23 depend on a parameter 2 (0 1]. So, we denote them by G13 , G23 and assume that

Denote by ah the bilinear form ah from (4.11.9) in which the terms u=R1 and v=R2 are omitted and G13 , G23 take the place of G13 and G23 . For the Timoshenko

! G13 , ! G23 are positive, continuous, decreasing9 = functions given on (0 1] such that G13 ! 1, G23 ! 1 as ! 0.

(4.11.17)

g satisfying

model, the problem of stress-strain state of the shallow shell is the following: Find

g = (u v w 1 ah (g g) = (q g) where q is a given element of V1 . !, is de ned as ~

2 ) 2 V1 g 2 V1

(4.11.18)

The solution of the corresponding model for the Kirchho model, denoted by

! = (~ v w) 2 V ~ u~ ~ ah (~ !) = (f !) ! !2V
where f 2 V and ah is determined by (4.7.8). De ne an operator P 2 L(V V1 ) as follows:

(4.11.19)

V 3 ! = (u v w) ! P! = u v w ; @w ; @w : @x @y

(4.11.20)

We suppose that the functionals of loading acting on the shell for the Timoshenko model q 2 V1 and for the Kirchho model f 2 V are equal in the following sense:

(f !) = (q P!) ! 2 V: (4.11.21) Theorem 4.11.2 Suppose that the conditions of Theorems 4.7.1 and 4.11.1 hold. Let f 2 V , q 2 V1 and let (4.11.21) be satis ed. Let also (4.11.17) be valid. Then

g ! P ! in (W21 ( ))5 as ! 0: ~

(4.11.22)

4.11. Finite shear model of a shell

283

Proof. 1) It follows from (4.11.18) that

Let
0 2 (0

ah (g g ) kqkV1 kg kV1 :
1). By (4.11.17) and Theorem 4.11.1 we get

(4.11.23)

ah (! !) ah 0 (! !) ck!k2 1 V

! 2 V1

2 (0 0 ]

(4.11.24)

where c is a positive constant. The inequalities (4.11.23) and (4.11.24) give

kg kV1 c;1 kqkV1 2 (0 0 ]: (4.11.25) Therefore, there exists a sequence f i g1 (0 0 ], lim i = 0, such that i=1 g i ! g0 weakly in V1 : (4.11.26)
From (4.11.9), (4.11.23), and (4.11.25) we obtain

G13 kP4 g k2 2( ) + G23 kP5 g k2 2 ( L L

c1

2 (0 0 ]:

(4.11.27) (4.11.28) (4.11.29)

This inequality together with (4.11.17) yields

P4 g i ! 0 and P5 g i ! 0 in L2 ( ) as i ! 1:
By de nition we have

P4 g i =

i1 +

@w i @x

P5 g i =

i2 +

@w i @y

and (4.11.26), (4.11.28) imply

g0 = (u0 v0 w0 @w0 01 = ; dx
Therefore, w0 2 W22 ( ) and Using (4.11.20), we get

01 02 ) 2 V1 @w0 02 = ; @y :

(4.11.30) (4.11.31)

!0 = (u0 v0 w0 ) 2 V

P!0 = g0 :

P4 P! = 0 P5 P! = 0 ! 2 V: (4.11.32) We accept in (4.11.18) = i , g = P!, where ! 2 V . Then, taking into account

(4.11.21), (4.11.26), (4.7.8), (4.11.9), and (4.11.32), we pass to the limit in (4.11.18) as i ! 1. As a result, we get

ah (!0 !) = (f !)

! 2 V:

(4.11.33)

284

Chapter 4. Direct problems for plates and shells

From the uniqueness of the solution of the problem (4.11.19), we obtain !0 = !, ~ and (4.11.26) is ampli ed in the sense

g ! g0 = P ! weakly in V1 as ! 0: ~

(4.11.34)

2) Let as before 2 (0 1). By Theorem 4.11.1, the bilinear form ah 0 generates the norm in V1 ; kgk1 = ah 0 (g g) 1=2 which is equivalent to the norm (4.11.16). Let us prove that lim0 ah 0 (g g ) = ah 0 (g0 g0) ! then upon (4.11.34) we will get (4.11.22). De ne a functional (g) = ah (g g) ; 2(q g) (g ) = g2V min 1 Therefore, (g ):
0 ]:

(4.11.35)

g 2 V1 :

The function g is a solution of the problem (4.11.18) if and only if

ah (g g ) ; 2(q g ) ah (g0 g0 ) ; 2(q g0 ) 2 (0 Since g0 = P !, ! 2 V , we obtain by (4.11.32) that ~ ~ ah (g0 g0 ) = ah 0 (g0 g0) 2 (0 0 ]:

(4.11.36) (4.11.37)

Passing to the limit in the left-hand side of (4.11.36), we conclude via (4.11.34) and (4.11.37) that lim0 sup ah (g g ) ah 0 (g0 g0): ! On the other hand, (4.11.17) gives The two last inequalities yield lim0 sup ah 0 (g g ) ah 0 (g0 g0 ): ! It follows from (4.11.34) that lim0 inf ah 0 (g g ) ah 0 (g0 g0 ): ! (4.11.39) Now, (4.11.38) and (4.11.39) give (4.11.35), which concludes the proof. (4.11.38)

ah (g g ) ah 0 (g g )

2 (0 0 ]:

4.12. Laminated shells

285

Remark 4.11.2 Let

I = ah (g g ) ; 2(q g ) I0 = ah (~ !) ; 2(f !) !~ ~ I
and (4.11.40) where g and ! are the solutions of the problems (4.11.18) and (4.11.19), respec~ tively. Then,

I0

(4.11.41)

G13 kP4 g k2 2( ) + G23 kP5 g k2 2( ) ! 0 as ! 0: (4.11.42) L L Indeed, since ah (g0 g0) = ah (~ !), see (4.11.34), we conclude from (4.11.21) and !~
(4.11.36) that (4.11.41) holds. Denote by a the bilinear form ah de ned by (4.11.9) with G13 = G23 = 0. We have

lim0 I = I0 !

a(g g ) ; 2(q g ):

(4.11.43) (4.11.44)

It follows from (4.11.21) and (4.11.22) that lim0 a(g g ) ; 2(q g )] = ah (~ !) ; 2(f !) = I0 : !~ ~ ! Now, by (4.11.41), (4.11.43), and (4.11.44), we get (4.11.42).

4.12 Laminated shells

We consider a composite, laminated, thin shell fabricated of an arbitrary number of homogeneous, anisotropic laminae. We suppose that the thickness of each lamina is constant, and so the thickness of the shell is constant. Let also the coordinate surface = 0 coincide with the midsurface of the shell, see Fig. 4.12.1 . We apply the Kirchho hypotheses for the whole stack of laminae, see Section 4.5 and Ambartsumian (1974). Let ! = (u v w) be the vector function of displacements of points of the surface = 0, u, v, w being the displacements in the directions of the coordinate lines , , . Denote by u , v , w the displacements of points situated at distance from the surface = 0. We have (see (4.5.3))

4.12.1 The strain energy of a laminated shell

u = 1 + R u ; A @w v = 1 + R v ; B @w @ @ 1 2 w = w:

(4.12.1)

286

Chapter 4. Direct problems for plates and shells

Here, A, B are the coe cients of the rst quadratic form, R1 , R2 the radii of the principal curvatures, the coordinate lines = const, = const coincide with lines of the principal curvatures of the surface = 0. The strain components are de ned by the formulas (4.5.4), (4.5.5) and the stress components given by 1 = E11 "1 + E12 "2 + E16 "12 (4.12.2) 2 = E21 "1 + E22 "2 + E26 "12 Eij = Eji : 12 = E61 "1 + E62 "2 + E66 "12 Here, Eij are elasticity coe cients, which in this case are (see Fig. 4.12.1) n Eij ( ) = Eji ( ) = Eij

2 ;h+ 2
=1 2

n;1

;h + 2

n Eij 's being constants. The strain energy (see (4.5.7) and (4.5.8)) is determined by
ZZ

0=0

i j = 1 2 6 n = 1 2 ::: s s=h
12 "12

(4.12.3)

d d

Z h ; 2
;h 2

1 "1

+ 2 "2 +

AB d :

(4.12.4)

In the expressions for H1 and H2 in (4.5.8), we neglect the terms R1 and R2 since they are small as compared to 1.
g

h/2
g=0
d d
1 2

Figure 4.12.1: Laminated shell We substitute (4.12.2) into (4.12.4). By (4.5.4) and (4.12.3), we obtain 1 ZZ ;c "2 + 2c " " + c "2 + c "2 + 2c " " + 2c " " AB d d =2 11 1 12 1 2 22 2 66 12 16 12 1 26 12 2 +
ZZ

K11 "1 1 + K12 ("1 2 + "2 1 ) + K22 "2 2 + 2K66 "12

ds d s =-1 h

12

4.12. Laminated shells

287

+ K16 (2"1 12 + "12 1 ) + K26 (2"2 12 + "12 2 ) AB d d 1 Z Z ;D 2 + 2D 2 2 +2 11 1 12 1 2 + D22 2 + 4D66 12 + 4D16 where
s X n Eij ( n ; n;1 ) n=1 s X n 2 2 Kij = 1 Eij ( n ; n;1 ) ; s ( n ; n;1 ) 2 n=1 s X n 3 3 2 3 2 2 Dij = 1 Eij ( n ; n;1 ) ; 3 s ( n ; n;1 ) + 4 s ( n ; n;1 ) 3 n=1 2
1 12 + 4D26 2 12

AB d d

(4.12.5)

cij =

(4.12.6)

4.12.2 Shell of revolution

For a shell of revolution, is de ned by (4.6.1) and the strain components are determined by (4.6.3){(4.6.5). We suppose that (4.6.15), (4.6.16) hold. Let
f f f W = W21 ( ) W21 ( ) W22 ( )

(4.12.7)

f where W2m ( ), m = 1 2, is the subspace of W2m ( ) consisting of periodic functions with respect to ', see subsec. 4.6.1. We denote by ! = (u v w) an element of W . De ne operators Pi 2 L(W L2 ( )) by (4.6.5), (4.6.10), and associate with the strain energy the following bilinear form on W :

a (! ! ) = c11 (P1 !0 )(P1 !00 ) + c12 (P1 !0 )(P2 !00 ) + (P2 !0 )(P1 !00 ) 0 0 + c22 (P2 !0 )(P2 !00 ) + c66 (P3 !0 )(P3 !00 ) + c16 (P1 !0 )(P3 !00 ) + (P3 !0 )(P1 !00 ) o + c26 (P2 !0 )(P3 !00 ) + (P3 !0 )(P2 !00 ) AB dz d'
0 00

Z LZ 2 n

K11 (P1 !0 )(P4 !00 ) + (P4 !0 )(P1 !00 ) 0 0 + K12 (P1 !0 )(P5 !00 ) + (P5 !0 )(P1 !00 ) + (P2 !0 )(P4 !00 ) + (P4 !0 )(P2 !00 ) + K22 (P2 !0 )(P5 !00 ) + (P5 !0 )(P2 !00 ) + 2K66 (P3 !0 )(P6 !00 ) + (P6 !0 )(P3 !00 ) + K16 2(P1 !0 )(P6 !00 ) + 2(P6 !0 )(P1 !00 ) + (P3 !0 )(P4 !00 ) + (P4 !0 )(P3 !00 ) o + K26 2(P2 !0 )(P6 !00 )+2(P6 !0 )(P2 !00 )+(P3 !0 )(P5 !00 )+(P5 !0 )(P3 !00 ) AB dz d'
+

Z LZ 2

288 +
Z LZ 2
0 0

Chapter 4. Direct problems for plates and shells

D11 (P4 !0 )(P4 !00 ) + D12 (P4 !0 )(P5 !00 ) + (P5 !0 )(P4 !00 ) + D22 (P5 !0)(P5 !00 ) + 4D66 (P6 !0 )(P6 !00 ) + 2D16 (P4 !0 )(P6 !00 ) + (P6 !0)(P4 !00 ) + 2D26 (P5 !0 )(P6 !00 ) + (P6 !0 )(P5 !00 ) AB dz d': (4.12.8)
Here, !0 !00 2 W and we denote the bilinear form by a since the coe cients cij , Kij , Dij depend on the vector = ( 1 : : : s ), see (4.12.6). It follows from (4.6.10), (4.12.5), and (4.12.8) that

a (! !) = 2 (!)

!2W

(4.12.9)

where (!) is the strain energy for a displacement !. We suppose that the elasticity coe cients satisfy the following conditions
X

i j =1 2 6

n 2 2 2 Eij j i c( 1 + 2 + 6 )

1 2 6) 2 R

n n n = 1 2 : : : s Eij = Eji

c = const > 0:

(4.12.10)

Theorem 4.12.1 Let the conditions (4.6.15), (4.6.16), (4.12.10) hold. Let V be a subspace of W satisfying (4.6.11). Then, the bilinear form a determined by (4.6.3){(4.6.5), (4.6.10), (4.12.6), and (4.12.8) is symmetric, continuous, and coercive in V , i.e.,
a (!0 !00 ) = a (!00 !0 ) !0 !00 2 V 0 00 0 00 a (! ! ) c1 k! kV k! kV !0 !00 2 V a (! !) c2 k!k2 ! 2 V: V
(4.12.11) (4.12.12) (4.12.13)

Here, c1 , c2 are positive constants, c2 depending on the constant c from (4.12.10) and on s .

Proof. By (4.5.4) and (4.6.10), we have

"1 (!) = P1 ! + P4 ! "2 (!) = P2 ! + P5 ! "12 (!) = P3 ! + 2 P6!
(4.12.14)

Taking into account (4.12.2){(4.12.4), (4.12.9), (4.12.10), (4.12.14), and denoting "6 (!) = "12 (!), we obtain

4.12. Laminated shells

Z LZ 2 Z h 2
;h 2

289

a (! ! ) =

0 0 Z LZ 2

AB dz d'

c3

0 0 Z LZ 2 X 6 0 0

AB dz d'
i=1

Z h 2

i j =1 2 6

Eij "i (!)"j (!) d

;h

("1 (!))2 + ("2 (!))2 + ("12 (!))2 d

(Pi !)2 dz d': (4.12.15)

This estimate together with (4.6.20) yields (4.12.13). The relations (4.12.11) and (4.12.12) are obvious.

4.12.3 Shallow shells

For these shells, the midsurface is determined by (4.7.1), the bilinear form a is de ned by (4.7.2), (4.7.3), (4.7.6), and (4.12.8), with A = B = 1, dz = dx, d' = dy, and the integration over (0 L) (0 2 ) is replaced with the integration over , where is an arbitrary bounded plane domain with a Lipschitz continuous boundary V is de ned by (4.7.5) and (4.7.7). By analogy with the above, we prove the following

Theorem 4.12.2 Suppose (4.7.9) and (4.12.10) hold. Then, the bilinear form a

for the shallow shell is symmetric, continuous, and coercive in V , i.e., (4.12.11){ (4.12.13) hold.

290

Chapter 4. Direct problems for plates and shells

Chapter 5

Optimization of deformable solids

\ \I thought I'd try and nd my way to the top of the hill|" said Alice. \When you say `hill,' " the Queen interrupted, \I could show you hills, in comparison with which you'd call that a valley." " | L. Carroll \Through the Looking Glass"

5.1 Settings of optimization problems for plates and shells

5.1.1 Goal functional and a function of control
Plates and shells are main elements of many advanced structures. One of the most important characteristics of a construction is its weight, which determines the consumption of the material needed for the production of the construction as well as some operating features of the latter. For example, the increase of the weight of an aircraft causes the growth of the fuel rate in ight and the degradation of some ight characteristics. The weight of a homogeneous plate is determined by the relation

J=

ZZ

h dx dy

(5:1:1) is the domain occupied by the

where is the speci c gravity of the material, 291

292

Chapter 5. Optimization of deformable solids

midplane of the plate, and h is the function of thickness. The weight of a threelayered plate is de ned by the following formula (see Fig. 4.3.1)

J=

ZZ

( 1 t0 + 2 2 h) dx dy

(5:1:2)

where 1 and 2 are the speci c gravities of the interior and exterior layers, respectively. The weight of a shell is given by the formula

J=

ZZZ

d d

Z h=2
;h=2

H1 H2 d :

(5:1:3)

Here, is the speci c gravity of the material, h is the function of thickness of the shell, and are orthogonal curvilinear coordinates, H1 and H2 are the Lame coe cients. Because of its great importance, in many problems of optimization of plates and shells, the weight either is considered as a goal functional, or is included into restrictions (i.e., into the set of admissible controls). Important characteristics of a structure (plate or shell) are the frequencies of the natural oscillations and the critical load under which the construction loses stability. These characteristics can de ne not only the set of admissible controls, but also the goal functional. For example, studied are problems in which one maximizes the rst natural frequency, the critical load under which the construction loses stability, etc. In some problems, one considers a goal functional corresponding to optimization by some nite number of gures of merit. For example, if the goal functional is of the form J (h) = c1 J1 (h) ; c2 J2 (h) where h is a control, J1 (h) is the weight, J2 (h) is the rst natural frequency, c1 and c2 are positive constants, then the optimization problem on nding the minimum of J (h) on Uad means the minimization of the weight and the maximization of the

rst natural frequency. As control functions in optimization problems for plates and shells, one can take the function of thickness, as well as the function determining the midsurface of the shell (see the formulas (4.5.1) and (4.6.2)). Sometimes, one creates special control loads so that the external load acting on the plate or shell can also be considered as a control. The set of admissible controls consists of a system of restrictions, which will be considered now.

5.1. Settings of optimization problems for plates and shells

293

5.1.2 Restrictions

Restrictions on geometry and eigenvalues

Since the thickness of a plate (shell) is bounded and positive, the function of thickness h must meet the estimates ^ h h h in ^ where h and h are positive constants. If the shape of the midsurface of the shell is a control, then one imposes restrictions on the corresponding control function. For example, for a shell of revolution, the function r(z ) (see the relations (4.6.2) and Fig. 4.6.1) can be a control. In this case, the condition

should be smooth in a sense (see Theorem 4.6.1). Restrictions on frequencies of the natural oscillation of the plate (shell) can also be imposed. In some cases, one demands that the rst natural frequency be not less than a xed value in other cases one wishes that neighborhoods of some xed numbers do not contain the frequencies of natural oscillations, etc. If the plate (shell) can lose stability under the action of a given load, then one considers a restriction on stability.

r1 r(z ) r2 z 2 0 L] must be ful lled, r1 and r2 being positive numbers. Moreover, the function r(z )

Restrictions on sti ness and strength

One considers restrictions on the vector function ! = (u v w) of displacements of points of the midsurface of the shell. Since u v 2 W21 ( ) and w 2 W22 ( ) (see Sections 4.6 and 4.7), the restrictions on u, v, w, e.g., for a shallow shell, may be taken in the form
(x y)2 ZZ
i

max jw(x y)j c

ZZ
i

u dx dy

ai
(5.1.4)

v dx dy

bi

i = 1 2 : : : k:

Here, i are subdomains of the domain such that mes i > 0, ai , bi , and c are positive numbers. In view of the inequalities (5.1.4), we stress that, since w 2 W22 ( ) C ( ), the displacements w can be controlled at every point, whereas the functions of displacements u and v do not belong, in general, to W22 ( ), so that they are controlled in the integral sense. Restrictions on strength are also of great importance. There exist a number of di erent strength criteria applicable to various materials (see Goldenblat

294

Chapter 5. Optimization of deformable solids

and Kopnov (1968), Wu (1974)). A generalized strength criterion for anysotropic materials has the form
X

ik

ik ik +

pqmn

pqmn pq mn

rstlmn

rstlmn rs tl mn +

; 1 0: (5.1.5)

Here, ik are the components of the stress tensor ik , pqmn , etc. are the components of the strength tensors of di erent valences. For a number of anysotropic materials, in the criterion (5.1.5) one can consider only the rst two summands, i.e., the strength criterion looks like
X

ik

ik ik +

pqmn

pqmn pq mn ; 1

0:

(5:1:6)

The components of the stress tensor for the shell are determined by the formulas (4.5.4){(4.5.6), and in this case 11 = 1 22 = 2 21 = 12 = 12 (5.1.7) 13 = 31 = 23 = 32 = 33 = 0: Restrictions on the strength must be ful lled at every point of the three-dimensional domain Q occupied by the plate or shell. Results of Chapter 4 imply that stresses in the plate (shell) are elements of the space L2 (Q). Thus, in order that the relations (5.1.5) or (5.1.6) make sense at every point ( ) 2 Q (see Fig. 4.5.1), one has to average (regularize) the stresses (see subsec. 1.6.4). To this end, the function of stresses must be extended to a larger domain. ( Let ij ) be an averaging of the function ij relatively the and coordinates, ( being the radius of the averaging kernel. ij ) are obviously functions of a point ( ( ( ) 2 Q, i.e., ij ) = ij ) ( ). According to the Kirchho hypotheses, the stresses along the normal to the midsurface of the shell change by a ne law (see the formulas (4.5.4){(4.5.6)), and so the conditions (5.1.5) and (5.1.6) are to be veri ed only on the surfaces of the plate (shell), i.e., for = h . Thus, e.g., for 2 the relation (5.1.6), the strength conditions take the form
~= 2

max h max h
;2

2 X

)2

i k=1

() ik ik ( 2 X

~)
( pqmn pq) ( () ~) mn (

p q m n=1

~) ; 1 0: (5.1.8)

We stress the operation of averaging (regularization) is legitimate because, on one hand, the relation (5.1.8) makes sense, and on the other hand, the very

5.2. Approximate solution of direct and optimization problems

:::

295

physical notion of a stress contains some averaging. In connection with this, one can use also integral restrictions on strength, i.e., demand that inequalities of the form (5.1.6) take place in the integral sense, on some small pieces of the surface of the shell. For example, for an isotropic shell, one can use the restrictions of the form
ZZ X 2 2 max h ik ( ~= h ; 2 2 j i k=1

~)AB d d ; cj

0 (5.1.9)

j = 1 2 : : : s cj = const > 0

A and B being coe cients of the rst quadratic form (see Section 4.5). Notice that
the domain is divided into the subdomains j .

5.2.1 Direct problems and spline functions

uh 2 V
(ui i ) 2 V

5.2 Approximate solution of direct and optimization problems for plates and shells
ah (uh v) = (f v) ui 6= 0 v2V v 2 V:
(5:2:1)

As shown in Chapter 4, the solution of problems on the stress-strain state of a plate or shell reduces to determining a function uh such that while the stability problems and the problems of nding frequencies of natural oscillation of plates and shells reduce to the eigenvalue problem

i ah (ui v ) = bh (ui v )

(5:2:2)

To get approximate solutions of the problems (5.2.1), (5.2.2), one can use the Riesz and Galerkin methods (interior approximation methods) as well as the exterior approximation methods, disturbed approximations, and other methods (see, e.g., Aubin (1972)). In subsecs. 2.4.4 and 2.6.3, we considered the Riesz and Galerkin methods of approximate solution of the problems (5.2.1), (5.2.2) and established that, if a sequence of nite-dimensional subspaces fVm g1=1 V satis es the condition m lim 2V m!1 uinfm ku ; v kV = 0

v2V

(5:2:3)

then the Riesz and Galerkin methods ensure the convergence of approximate solutions to the explicit ones. Speci c forms of the space V for di erent models of plates and shells were presented in Chapter 4. Below, we cite few examples of the sequences fVm g1=1 m V meeting the condition (5.2.3).

296

Chapter 5. Optimization of deformable solids

In what follows, we will need the notion of a spline. Let there be de ned a partition of an interval a b]: : a = x0 < x1 < < xN = b: (5:2:4) By Sp(m) ( a b]), m 2 N , we denote the space consisting of real-valued functions w such that w 2 C 2m;2 a b] and on each interval xi xi+1 ] the function w is a polynomial of degree 2m ; 1. The spaces Sp(m) ( a b]) for m = 1, 2 are of special importance for applications. They are called a ne and cubic spline spaces, respectively. De ne the following subspaces of the space Sp(2) ( a b]): Sp(2) ( 1 Sp(2) ( 2

Sp(2) ( 5

a b]) = s j s 2 Sp(2) ( a b]) s(x0 ) = s(xN ) = 0 a b]) n o ds ds = s j s 2 Sp(2) ( a b]) s(x0 ) = s(xN ) = dx (x0 ) = dx (xN ) = 0 Sp(2) ( a b]) 3 n o ds = s j s 2 Sp(2) ( a b]) s(x0 ) = s(xN ) = dx (x0 ) = 0 o n ds Sp(2) ( a b]) = s j s 2 Sp(2) ( a b]) s(x0 ) = dx (x0 ) = 0 4 Sp(2) ( a b]) 5 n o dp dp = s j s 2 Sp(2) ( a b]) dxs (x0 ) = dxs (xN ) p = 0 1 2 (5.2.5) p p a b]) is called the periodical spline space. Bases in the spaces Sp(2) ( a b]) and Sp(2) ( a b]) i = 1 2 ::: 5 i

are formed by corresponding fundamental splines, which are numerically constructed by using e ective algorithms (see Alberg et al. 1967, Laurent (1972), Zavialov et al. (1980)).

5.2.2 The spaces Vm for plates

We consider the model of a plate based on the Kirchho hypotheses, i.e., the Kirchho model. In this case, V W22 ( ). Suppose that is a rectangular domain, = (x y) j a1 < x < b1 a2 < y < b2 (5:2:6) ai and bi being constants. Let f n = 1n 2n g1 be a sequence of partitions n=1 of the rectangle = a1 b1 ] a2 b2 ], where < xNnn = b1 1n : a1 = x0n < x1n <

5.2. Approximate solution of direct and optimization problems

2n : a2 1n 2n

:::

297 (5.2.7)

We set

= y0n < y1n <

1n

< yMn n = b2 :
= 0 imin ;1 x(i+1)n ; xin N
n

= 0 imax ;1 y(i+1)n ; yin 2n = 0 imin ;1 y(i+1)n ; yin Mn Mn = max ( 1n 2n ) = min ( 1n 2n ) : n n n The notation n 2 P ( ) will denote that n , where > 0. Consider the case when the plate is clamped on the boundary. Then V = W 2 ( ) (see subsec. 4.1.4). De ne spaces Vn as follows 2 Vn = Sp(2) ( 1n a1 b1 ]) Sp(2) ( 2n a2 b2 ]) (5:2:8) 2 2 the symbol denoting the tensor product. Notice that Vn W 2 ( ). 2 2 ( ) \ W 1 ( ) (see If the plate is supported on the boundary, then V = W2 2 subsec. 4.1.4). De ne in this case the spaces Vn as Vn = Sp(2) ( 1n a1 b1 ]) Sp(2) ( 2n a2 b2 ]): (5:2:9) 1 1 Now, Vn W22 ( ) \ W 1 ( ). Using results on the interpolation of splines in a 2 rectangular domain (see Zavialov et al. (1980)) and taking into account that the set of smooth functions is dense in V , one can easily see that the spaces Vn de ned by the relations (5.2.8) and (5.2.9) for the clamped and supported plates, respectively, meet the conditions (5.2.3) provided that the corresponding sequence of partitions f n g is such that n 2 P ( ) for all n and n ! 0. In the same way, using the tensor product of the subspaces of the form (2) , Sp(2) , Sp(2) , and Sp(2) (see (5.2.5)), one can construct a sequence of spaces Sp1 2 3 4 satisfying (5.2.3) for the cases when, on di erent edges of the plate, one imposes di erent boundary conditions, such as those of clamping, supporting, and of free edge. A more complicated case is that when one considers di erent boundary conditions on the same edge. Then, one can use splines together with a method of exterior approximation (see Litvinov (1981a)). Now, let be a non-rectangular domain. If the plate is clamped or supported on the whole boundary, then one can use the \classic method" of construction of the spaces Vn , according to which these spaces are set up by using the elements of the form (see Michlin (1970))

= 0 imax ;1 x(i+1)n ; xin N

n

u=

n X i=1

ci !k 'i

(5:2:10)

where ci are constants, ! is a function which is positive in and vanishes on S , the boundary of , f'i g1 is some system of functions, k = 1 in the case of i=1 supporting and k = 2 for clamping.

298

Chapter 5. Optimization of deformable solids

Rvachev proposed an e ective method of construction of functions ! which satisfy the above conditions and given smoothness requirements for domains of a complicated shape (see Rvachev, V.L. and Rvachev, V.A. (1979)). To obtain approximate solutions of direct problems for plates of a nonrectangular form, one can use also disturbed approximations (see Aubin (1972)). The application of the nite element method in the case when V W22 ( ), is of a complicated form, and is divided into triangles, leads to great di culties (see, e,g., Ciarlet (1978)). ; 1 For the nite shear model of plates, V W2 ( ) 3 (see Section 4.11, for a bending of the plate the vector of displacements has the form ! = (w 1 2 )). ; The approximation in a subspace of the space W21 ( ) 3 is rather easy and e ectively realized by nite elements for domains of complicated forms. Hence, for such domains it is natural to use not the Kirchho model, but the more precise nite shear model, and to solve a corresponding problem by nite elements. (In this case, we faces the situation when a re nement of the model simpli es a corresponding problem in the above mentioned sense.)

5.2.3 The spaces Vm for shells

Shallow shell
For the Kirchho model, V is the subspace W21 ( ) W21 ( ) W22 ( ) (see Section 4.7). If is a rectangular domain determined by the relation (5.2.6) and on each edge of assigned is one of the following conditions: clamp, support, or free edge, then the spaces Vm V satisfying the condition (5.2.3) are constructed by splines just as it was done for the plates. For example, consider a shell clamped on the whole edge. Now, the space V has the form

V = ! = (u v w) j ! 2 W 1 ( ) W 1 ( ) W 2 ( ) : 2 2 2
(j ) 1 be a sequence of partitions of the rectangle Let (j) = (jn) n 1 2n n=1 that j = 1, 2, 3 and

(5:2:11) such

(j ) 1n (j ) 2n

) (j = 0 imax ;1 y((ij+1)n ; yin) Mnj (j ) (j ) (j ) n = max 1n 2n

= 0 imax ;1 N
nj

) : a1 = x(jn) < x(jn) < < x(jnj n = b1 0 1 N ( ( (j : a2 = y0j) < y1jn) < < yM)nj n = b2 n ) (j ) = min (j ) (j ) x(j+1)n ; x(j) 1n 0 i N ;1 x(i+1)n ; xin in (i (j ) 2n (j ) ) (j = 0 i min ;1 y((ij+1)n ; yin) Mnj (j ) 1n (j ) 2n
nj

(j ) 1n (j ) 2n

n = min

(5.2.12)

5.2. Approximate solution of direct and optimization problems

:::

299

We suppose that

(j ) g1 n n=1

P( )

lim (j) = 0 n!1 n

j = 1 2 3:

(5:2:13) (5:2:14)

De ne the spaces Vn in the following way:

Vn = Vn(1) Vn(2) Vn(3)

where

) ) Vn(i) = Sp(2) ( (in a1 b1 ]) Sp(2) ( (in a2 b2 ]) i=1 2 1 1 1 2 (3) = Sp(2) ( (3) a b ]) Sp(2) ( (3) a b ]): Vn (5.2.15) 2 1n 1 1 2 2n 2 2 In this case Vn V , where V is de ned by (5.2.11). Since the set of smooth functions is dense in V , making use of the error estimates of interpolation by cubic splines of two variables (see Zavialov et al. (1980)), one can show that the spaces Vn

de ned by the formulas (5.2.14) and (5.2.15) satisfy the condition (5.2.3) provided (5.2.13) holds. We point out that the spaces Vn(1) and Vn(2) could be constructed as tensor products of subspaces of those a ne splines that vanish at the points ai and bi . (3) If Vn is determined by the relation (5.2.15), the inclusion Vn V holds, and provided (5.2.13) is true, (5.2.3) is valid. If the shell has a non-rectangular plan, i.e., is a non-rectangular domain, then in order to construct the spaces Vn one can use the above stated methods for plates. Just as in the case of a plate, if has a rather complicated form, it ; 1 is useful to consider a nite shear model (for this model V W2 ( ) 5 ) and to construct spaces Vn on the basis of nite elements.

Shell of revolution

For a shell of revolution, is a rectangular domain de ned by the relation = (z ') j 0 < z < L 0 < ' < 2 (5:2:16)
f f f f and V W21 ( ) W21 ( ) W22 ( ), where W2l ( ), l = 1 2, are the spaces of periodical functions (see Section 4.6). For a clamped shell, the space V has the form f f f V = ! = (u v w) j ! 2 W21 ( ) W21 ( ) W22 ( ) o u S1 = v S1 = w S1 = @w S = 0 @z 1 n

(5.2.17) (5:2:18)

and

S1 = (z ') j z = 0 L 0 < ' < 2

300

Chapter 5. Optimization of deformable solids

If the shell is supported, then f f f V = ! = (u v w) j ! 2 W21 ( ) W21 ( ) W22 ( ) u S1 = v S1 = w S1 = 0 : (5.2.19) Let f (j) = (jn) (jn) g1 be the sequence of the partitions of the rectangle n 1 2 n=1 , j = 1 2 3, determined by the relations (5.2.12) provided z and ' are substituted for x and y, a1 = 0, b1 = L, a2 = 0, b2 = 2 . For these partitions, the condition (5.2.13) is supposed to be valid. The spaces Vn are of the form Vn = Vn(1) Vn(2) Vn(3) (5:2:20) where ) ) Vn(i) = Sp(2) ( (in 0 L]) Sp(2) ( (in 0 2 ]) i=1 2 1 1 5 2 (3) = Sp(2) ( (3) 0 L]) Sp(2) ( (3) 0 2 ]) Vn 2 1n 5 2n if the shell is clamped, and

i=1 2 3 if the shell is supported. Notice that, in both cases, the inclusion Vn

V takes place and provided (5.2.13) is valid, (5.2.3) is true. Analogously to the above, in the form of tensor products of corresponding spline spaces, one can construct spaces Vn V satisfying the condition (5.2.3) in the case when one edge is clamped while the other edge is either supported or free. Further, let f (i) g1 be a sequence of partitions of the segment 0 L]: n n=1
(i) (i) n : 0 = z0n < z1n < (i)

) ) Vn(i) = Sp(2) ( (in 0 L]) Sp(2) ( (in 0 2 ]) 1 1 5 2

and

( < zNi) n = L ni

i=1 2 3
(5:2:21)

Denote by Pk the vector space of trigonometric polynomials of degree k, i.e., the totality of elements of the form = a0 +
k X j =1

(i) n = 0 jmax ;1 Nni

(i z((ji)+1)n ; zjn) :

(bj cos j' + cj sin j')

(5:2:22)

a0 , bj , and cj being constants.

For a shell of revolution, the spaces Vn can be constructed as tensor products of spline and trigonometric polynomial spaces. Then, the spaces Vn are of form (5.2.20), and for a clamped shell Vn(i) are determined by

Vn(1) = Sp(2) ( (1) 0 L]) PMn (1) n 1

5.2. Approximate solution of direct and optimization problems

:::

301 (5.2.23) (5:2:24)

and for a supported shell Vn(i) looks like

Vn(2) = Sp(2) ( (2) 0 L]) PMn (2) n 1 (3) = Sp(2) ( (3) 0 L]) P (3) Vn n 2 Mn i = 1 2 3:

The spaces Vn de ned by the formulas (5.2.20), (5.2.23), and (5.2.20), (5.2.24) for the clamped and supported shells, respectively, meet the condition of inclusion Vn V . Then, making use of results by Litvinov (1981b), one can easily see that ( these spaces satisfy the condition (5.2.3) provided Mni) ! 1 as n ! 1 and (i) = 0. limn!1 n As tensor products of spline and trigonometric polynomial subspaces, one can construct spaces fVn g V satisfying the condition (5.2.3) for other cases of xing of a shell for example, for the cases when one edge is clamped and the other edge is supported or free. Calculations have shown a high e ciency of the use of the subspaces of the form (5.2.23), (5.2.24), when the thickness of the shell h varies only along the generatrix, i.e., h = h(z ).

Vn(i) = Sp(2) ( (i) 0 L]) PMni) ( n 1

5.2.4 Direct problems for nonfastened plates and shells

Above, we always considered problems in which a bilinear form ah is coercive in V , i.e., ah (! !) ck!k2 !2V (5.2.25) V
Q

hold. Let us present an approach to the solution of such problems, which is based upon the work by Krizek and Litvinov (1992). From the theorems on the coerciveness of the forms ah for plates and shells, it follows that there exists a positive constant c1 such that ah (! !) + k!k2 2( )k c1 k!k2 ! 2 W: (5.2.26) W L Denote by W1 the kernel space of the form ah : W1 = ! j ! 2 W ah (! !0) = 0 !0 2 W : (5.2.27) W1 is a nite-dimensional subspace of functions with zero point strain energy, see Section 4.6.3. In the theory of elasticity, W1 is known as the subspace of rigid displacements. In the case when R2 , it is de ned by (1.7.25). For the Kirchho model of plate, W1 is the three-dimensional subspace with basis 1, x, y. For the models of shells, theses subspaces are given by the relations (4.6.22) and (4.7.10).

V being a subspace of the space W = k=1 W2li ( ). However, in case of a nonfasi tened plate (shell), V is not a subspace of W , but V = W and (5.2.25) does not

302

Chapter 5. Optimization of deformable solids

Let f'i gN be a basis in W1 and let W2 be the orthogonal complement of i=1 W1 to W with respect to the scalar product in L2( )N , i.e., W2 = ! j ! 2 W (! 'i )L2 ( )N = 0 i = 1 : : : N : (5.2.28) According to (5.2.26) and Theorem1.7.3, there exists a constant c1 such that ah (! !) c1 k!k2 ! 2 W2 : (5.2.29) W The problem on stress-strain state of a nonfastened plate (shell) consists in nding a function ! such that ~

!2W ~
where

ah (~ !) = (f !) !
(f ! ) = 0

!2W ! 2 W1 :

(5.2.30) (5.2.31)

f2W

The condition (5.2.31) is necessary and su cient for the solvability of the problem (5.2.30). The solution of the latter problem is de ned up to an arbitrary function of W1 , i.e., if ! is a solution of the problem (5.2.30), then !1 = ! + !, where ! 2 W1 , ~ ~ ^ ^ is also a solution of this problem. Therefore, it is an inconvenient problem from the point of view of computation. Instead of (5.2.30), we consider the following problem: Find a function ! satisfying

!2W

ah (! !) + r

N X i=1

(! 'i )L2 (

)N (!

'i )L2 (

)N

= (f ! )

!2W
(5.2.32)

where r is a positive constant. By virtue of Theorem 1.7.2 , the bilinear form ah de ned by

ah (!0 !00 ) = ah (!0 !00 ) + r

N X i=1

(!0 'i )L2 (

)N (!

00

' i )L 2 (

)N

!0 !00 2 W

is coercive in W for an arbitrary r > 0. So, there exists a unique solution ! of the problem (5.2.32). One can verify that the function ! is also a solution of the problem (5.2.30), i.e., ! = !, and in addition ! 2 W2 . Thus, the solution ! of the ~ problem (5.2.32) is independent of r.

5.2.5 Solution of optimization problems

Hn H

Let H be a Banach space of controls and let fHn g1 be a sequence of niten=1 dimensional subspaces such that

8n

(5.2.33)

5.2. Approximate solution of direct and optimization problems

nlim uinf n ku ; v kH = 0 !1 2H

:::

303 (5.2.34)

v 2 H:

~ The nite-dimensional optimization problem consists in nding a function h such that ~ ~ h 2 Hn \ Q f (h) = inf f (h) (5:2:35)

Q being a closed set in H .

h2Hn \Q

In particular, if the control is the function of the thickness of the plate (shell), then H = Wp1 ( ), with p > 2. In this case, the spaces Hn satisfying the conditions (5.2.33) and (5.2.34) can be constructed either as tensor products of onedimensional spline spaces, or by using the nite element method. From the point of view of numerical realization, the spaces

and 5.2.2). If is a polygon, it is convenient to divide it into triangles and to choose Hn as the space of continuous functions on that are a ne in each triangle, that is, to construct Hn on basis of Courant triangles (see, e.g., Ciarlet (1978)). However, if is a non-rectangular domain and

Hn = Sp(1) ( 1n a1 b1 ]) Sp(1) ( 2n a2 b2 ]) (5:2:36) are the most convenient if = a1 b1 ] a2 b2 ] (for the notations, see subsecs. 5.2.1

a1 b1]

a2 b2 ]

then Hn can be de ned as spaces of restrictions on of functions from the space de ned by (5.2.36). Let 1n and 2n be uniform partitions, i.e., in the relations (5.2.7)

xin = a1 + q1 i
where

yjn = a2 + q2 j

i = 0 1 : : : Nn j = 0 1 : : : Mn

Then, in the space Hn determined by (5.2.36), a basis is formed by the following fundamental splines
8 < Bij (x y) = : 1 ;

; q1 = b1N a1
n

; q2 = b2M a2 :
n

(5:2:37)

i ; x ; a1 q
1

1 ; j ; y ; a2 q
2

if (x y) 2 Rij (5:2:38) if (x y) 62 Rij

where

Rij = (x y) j (i ; 1)q1 x ; a1 (i + 1)q1 (j ; 1)q2 y ; a2 (j + 1)q2 i = 0 1 : : : Nn j = 0 1 : : : Mn : (5.2.39)

For a cylindric shell, when the control is the function of the thickness that varies only along the axial coordinate, the set of controls has the form H =

304

Chapter 5. Optimization of deformable solids

Wp1 ((0 L)), L being the length of the shell. The spaces Hn can be chosen as follows Hn = Sp(1) ( n 0 L])
where
n : 0 = z0 < z1 <

(5:2:40)

< zn = L:

(5:2:41)

If n = 0 max;1(zi+1 ; zi ) and n ! 0 as n ! 1, the conditions (5.2.33) and i n (5.2.34) hold. An arbitrary function f from the P Hn determined by the formulas space (5.2.40), (5.2.41), has the form f (z ) = n=0 fi 'i (z ), where fi = f (zi ) and 'i i are the functions of the form
8 z ; zi;1 > >z ; z > < i i;1 'i (z ) = > zi+1 ; z > zi+1 ; zi > :

if z 2 zi;1 zi ] if z 2 zi zi+1 ] if z 62 zi;1 zi+1 ]

8 < z1 ; z if z 2 z0 z1 ] '0 (z ) = : z1 ; z0 0 if z 62 z0 z1 ] 8 z;z n;1 if z 2 z < n;1 zn ] 'n (z ) = : zn ; zn;1 0 if z 62 zn;1 zn]:

i = 1 2 ::: n;1

(5.2.42)

After de ning a basis in Hn , (5.2.35) becomes a problem of nonlinear programming. Methods for solving such problems were set forth and studied by many authors see, e.g., Cea (1971), Himmelblau (1972), Minoux (1989). Below, we present some numerical solutions of optimization problems, in which as a control we consider the functions of the thickness ( rst problem) and the load (second problem). For the solution of corresponding problems of nonlinear programming, di erent methods were applied in order to reduce the problem with restrictions to one without restrictions: the method of penalty functions for the rst problem the method of duality and the method of augmented Lagrange multipliers for the second problem. For solution of minimization problems without restrictions, one used the conjugate gradient method as well as direct methods, which do not use derivatives.

5.3. Optimization problems for plates (control by the function of the thickness) 305

5.3.1 Optimization under restrictions on strength

Setting of the problem

5.3 Optimization problems for plates (control by the function of the thickness)

Consider the problem of optimization with respect to the weight of a plate of variable thickness, which is subject to a distributed load. Treating the function of the thickness of the plate as a control, de ne a set of admissible controls

Uad = h j h 2 Wp1 ( ) p > 2 khkWp1 ( ) c ^ h h h k (h uh ) 0 k = 1 2 : : : l : (5.3.1) Here, the positive numbers h and ^ are the lower and upper restrictions on the h thickness of the plate, uh is the solution of the problem on the stress-strain state
of the plate, i.e., (5:3:2) The load f is supposed to be given and xed. The bilinear form ah is determined by the relations (4.1.13) and (4.1.24) for isotropic and orthotropic plates, respectively, where D, Di , and D12 are de ned by the formulas (4.1.5) and (4.1.11), k (h uh) are the functionals of the restrictions on the strength which may be determined by the left hand sides of the inequalities (5.1.9) or (5.1.8) (in the latter case l = 1 in (5.3.1)), and for plates = x, = y, A = B = 1. These functionals satisfy the condition (2.2.20). As a goal functional we take the weight of the plate, which is given by the relation ZZ f1 (h) = h dx dy (5:3:3) being the speci c gravity of the material. Then, results of subsec. 2.2.2 imply that, if the set Uad is not empty, there exists a function h0 such that h0 2 Uad f1(h0 ) = h2U f1 (h): inf (5:3:4)
ad

uh 2 V

ah (uh v) = (f v)

v 2 V:

To get an approximate solution of the problem (5.3.4), one can use the technique of Sections 2.3 and 2.4.

Optimization of a square plate

As an example, let us consider the optimization problem for a square isotropic plate which is subject to uniformly distributed load. In this case, = (x y) j ; a x a ;a y a (5:3:5)

306

Chapter 5. Optimization of deformable solids

ZZ

the bilinear form ah is de ned by (4.1.13), and (f v) = C

v dx dy

(5:3:6)

where the constant C stands for the intensity of the distributed load. We examined two cases of the fastening of the plate, namely, clamping and supporting. For the clamp, the space V is of the form

n o @u V = u j u 2 W22 ( ) u x= a = u y= a = @x x= a = @u y= a = 0 (5:3:7) @y

and for the support

V = u j u 2 W22 ( ) u x= a = u y= a = 0 :

(5:3:8)

To construct approximate solutions of the problem (5.3.2), the spaces Vn were chosen in the form (5.2.8) for the clamp, and in the form (5.2.9) for the support. The symmetry allowed us to use the product of one-dimensional splines which are symmetric with respect to 0, i.e., we applied spaces of bicubic splines symmetric with respect to the x and y axes. To approximate the space Wp1 ( ), we used the bia ne spline spaces Hn of the form (5.2.36). Because of the symmetry in x and y, Hn were also constructed as tensor products of spaces of splines symmetric with respect to 0. The restrictions on the strength were taken on basis of the Norris strength criterion:
2 2 2 b 2 k (h uh ) = ~xx (h (x y )k ) + ~yy (h (x y )k ) + 2 ~xy (h (x y )k ) ; b 2

(5.3.9)

k = 1 2 : : : l:

2 2 f(x y)k glk=1 is a given set of points from . Further, ~xx (h (x y)k ), ~yy (h (x y)k ), 2 and ~xy (h (x y)k ) are the mean squared stresses on the area element k containing the point (x y)k : 2 ~xx (h (x y)k ) = S ( 1 ) k 2 ~yy (h (x y)k ) = S ( ) k 2 ~xy (h (x y)k ) = S ( ) k

Here, b and

are the tension and shear strength of the material of the plate,

ZZ
k ZZ k ZZ k

2(1 ; 2(1 ;

Eh

Eh

2(1 +

Eh

@ 2 uh @ 2 uh 2 ) @x2 + @y 2 @ 2 uh @ 2 uh 2 ) @y 2 + @x2 @ 2 uh 2 dx dy ) @x@y

2 2

dx dy dx dy
(5.3.10)

5.3. Optimization problems for plates (control by the function of the thickness) 307

where S ( k ) is the area of k E and are the elasticity modulus and Poisson's ratio of the material of the plate. The nite-dimensional optimization problem corresponding to (5.3.4) consists in nding a function hn such that

hn 2 Hn \ Uad

f1 (hn ) = h2Hinf U f1 (h) \

n
ad

(5:3:11)

f1 being de ned by the formula (5.3.3). hni 2 Hn

The problem (5.3.11) was solved by the method of penalty functions. So, we constructed a sequence fhni g1 such that i=1

f1i (hni ) = hinf f1i (h): 2H

n

(5:3:12) (5:3:13)

Here, f1i is the augmented functional, 1 f1i (h) = f1 (h) + r p(h)

i

where r1i is the penalty coe cient, ri > 0, ri ! 0 as i ! 1, and p(h) is the penalty function which has the form

p(h) = c1 khkWp1 ( ) ; c
ij

l X 2 + c2 ( k (h uh ))2 + + k=1 X X ^+ + c3 (h ; hij )2 + c4 (hij ; h)2 + ij

(

(5.3.14)

where

(e)+ = 0 if e < 0 e if e 0 ci are positive numbers, and hij = h (xin yjn ), i.e., hij are the values of the function h(x y) at the corresponding knots of the net. It can be shown that from the sequence fhni g1 one can choose a subsei=1 quence fhnj g1 such that hnj ! hn as j ! 1, where hn is a solution of the j =1 problem (5.3.11). We would like to stress that, in optimization problems for plates ^ and shells, the set Uad is often de ned illegibly. For example, the constants h, h, and c from (5.3.1) can take values from some domain. The same situation can be with the functionals k (h uh ), for instance, in (5.3.9) b and can run over some region. Therefore, there is no need to solve the problem (5.3.12) with very small ri . So, the main shortcoming of the method of penalty functions|the di culty of the solution of the problem of minimization of the augmented functional under a great penalty coe cient|does not arise in the solution of many optimization problems for plates and shells.

308

Chapter 5. Optimization of deformable solids

The functionals h ! k (h uh) are not convex, so there exists a whole set of local minima of the problems (5.3.11) and (5.3.12). (Here, by a local minimum ~ ~ of the problem (5.3.12) we mean a function h such that f1i (h) = inf h2Q f1i (h), ~ . Similarly, a function ~ is a Q being an open set in Hn containing the point h h ~ local minimum of the problem (5.3.11) if f1 (h) = inf h2Q\Uad f1 (h).) In view of this, nding the global minimum, i.e., the solution of the problems (5.3.11) and (5.3.12), is usually a rather complicated task, so that one would be content with a \good local minimum." To get the results cited below, the problem (5.3.12) was solved by the modied method of conjugate gradients, see Zangwill (1969). The local minima of the problem (5.3.12) were found for various initial approximations, the minimal one of them was determined to be an approximate solution of the problem (5.3.12). Calculations were carried out for square plates (see (5.3.5)) for the following data: E = 19:6 1010 Pa, = 0:3, b = 2:94 108 Pa, = 2:3 108 Pa, a = 5 cm, ^ h = 0:35 cm, h = 1:5 cm, C = 9:8 104 Pa, C being the intensity of the load (see (5.3.6)), and the partitions of for Hn were of the form
1n 2n

In view of the symmetry with respect to x, y, and the diagonal of the square, the dimension of the bia ne spline space Hn corresponding to the partition (5.3.15) was equal to 6. Below, in the table, cited are the relative values of the function of the thickness of the plate at the nodes for two essentially di erent optimal forms under clamped and supported edges. Way of fastening N1 Clamp I II Support I II Here,
ij =

: ;a = y0n < ; a = y1n < 0 = y2n < a = y3n < a = y4n : (5.3.15) 2 2

: ;a = x0n < ; a = x1n < 0 = x2n < a = x3n < a = x4n 2 2

0.35 0.35 1.35 1.49

22

0.97 0.35 0.49 1.51

23

0.61 0.35 0.78 0.76

24

0.72 1.24 0.95 0.80

33

1.00 1.24 1.25 0.80

34

0.70 1.31 1.06 1.16

44

N2
17 12 27 14

~ h is the function of the thickness of the optimal plate, h0 is the thickness of

~ (xin yjn ) h h0

the plate with constant thickness in which sresses reach the limit values, i.e., maxk k (h0 uh0 ) = 0 N1 is the variant number, and N2 is the gain in the weight (in per cents) as compared to the plate of thickness h0 . Figs. 5.3.1 and 5.3.2 show the distribution of the thickness in the optimal plates which correspond to the variant I for the clamp (Fig. 5.3.1) and for the

5.3. Optimization problems for plates (control by the function of the thickness) 309

support (Fig. 5.3.2), because of the symmetry, only a quarter of each plate is displayed.

Figure 5.3.1: Distribution of the thickness in the optimal plate for the clamp

Figure 5.3.2: Distribution of the thickness in the optimal plate for the support

Remark 5.3.1 As has been mentioned above, since the functionals h ! k (h uh) are not convex, there exists a whole set of the local minima of the problem (5.3.12). However, if the thickness of the plate varies within relatively small bounds, then optimization problems for such a plate can be reduced to the problem of minimization of a linear functional under linear restrictions, see Bratus (1981) In such a problem, every local minimum is the global one. Remark 5.3.2 In the above example, the restriction khkWp1(
) c was not used (see (5.3.1)), because for a nite-dimensional optimization problem it always holds provided c is large enough. However, if one solves the problem for various values of n, then in order to carry out the passage to the limit one has to take into account this restriction.

310

Chapter 5. Optimization of deformable solids

5.3.2 Stability optimization problem

Let be a bounded domain occupied by the midplane of a plate, and let S be the boundary of . Suppose that S = S1 S2 S3 , the sets Si are disjoint, the plate is clamped on S1 , and on S2 it is subject to a load Q ( > 0 being a parameter) that acts in the (x y) plane (see Fig. 5.3.3). Thus,

Deformation of the plate in the (x y) plane

Q = (Q1 Q2 Q3 )

Q1 Q2 2 L2 (S2 ) Q3 = 0
;
2

(5:3:16)

Q1 , Q2 and Q3 being the components of the vector function of load in the x, y, and z axes. Introduce a space V1 by V1 = g j g = (g1 g2 ) g 2 W21 ( )
;

g S1 = 0 :

(5:3:17)

V1 being equipped with the topology of the space W21 ( ) 2 is a Banach space.
W
S1 S2
x

l Q

Figure 5.3.3: Plate at the buckling The strain energy of the plate connected with deformations in the (x y) plane generates the following bilinear form on V1 V1 (see Section 4.2)
ZZ

h E11 @q1 @g1 + E22 @q2 @g2 + E12 @q1 @g2 + @q2 @g1 @x @x @y @y @x @y @y @x (5.3.18) + G @q1 + @q2 @g1 + @g2 dx dy @y @x @y @x q = (q1 q2 ) 2 V1 g = (g1 g2 ) 2 V1 : Here, h is the function of the thickness of the plate, Eij and G are the elasticity a(1) (q g) = h
constants:

Eii = 1 ;Ei

1 2

i=1 2

E12 = E21 = 2 E11 = 1 E22 :

(5:3:19)

5.3. Optimization problems for plates (control by the function of the thickness) 311

For all h 2 Yp , Yp being de ned by the relations (2.1.2) and (2.1.3), the bilinear form a(1) is symmetric, continuous, and coercive on V1 V1 (see Lemma h 4.2.1), and hence because of (5.3.16) there exists a unique function qh such that

qh = (qh1 qh2 ) 2 V1

a(1) (q
h

h g) =

S2

(Q1 g1 + Q2 g2 ) ds

g 2 V1 : (5:3:20)

Here, qh1 and qh2 are the components of the vector function of displacements in the x and y axes of points of the midplane of the plate in the case = 1. These displacements cause the following forces acting in the midplane of the plate:

N1 (qh ) = E11 h @qh1 + 2 @qh2 @x @y @qh2 + @qh1 N2 (qh ) = E22 h @y 1 @x N3 (qh ) = Gh @qh1 + @qh2 : @y @x

(5.3.21)

Lemma 5.3.1 Let the conditions (4.1.25){(4.1.27) hold, and let a set Yp be de ned

by the relations (2.1.2) and (2.1.3), p > 2. Assume that a form a(1) is determined by h the formulas (5.3.18) and (5.3.19), and that (5.3.16) is valid. Then, the functions h ! Ni (qh ), i = 1 2 3, de ned by (5.3.20) and (5.3.21) are continuous mappings of Yp equipped with the topology generated by the Wp1 ( )-weak topology into L2 ( ):

From here, just as it was done in the proof of Lemma 2.10.1, we conclude that

Proof. Let fhng1=1 Yp and hn ! h0 weakly in Wp1( ). Then, by the imbedding n theorem, we have hn ! h0 in C ( ): (5:3:22)
qhn ! qh0 Ni (qhn ) ! Ni (qh0 )
in V1 (5:3:23)

where qhn is the solution of the problem (5.3.20) for h = hn , n = 0 1 2 : : : Now, (5.3.21){(5.3.23) imply that in L2 ( ) as n ! 1

i=1 2 3

concluding the proof.

Direct stability problem

Let
o n (5:3:24) V2 = f j f 2 W22 ( ) f S1 = 0 @f S = 0 @ 1 @f being the derivative with respect to the normal to S1 . The set V2 being endowed @ with the norm of the space W22 ( ) is a Banach space.

312

Chapter 5. Optimization of deformable solids

V2 V2

The bending energy of the plate generates the following bilinear form on

a(2) (w0 w00 ) = h

+ E12

h3 E @ 2 w0 @ 2 w00 + E @ 2 w0 @ 2 w00 22 @y 2 @y 2 12 11 @x2 @x2 @ 2 w0 @ 2 w00 + @ 2 w0 @ 2 w00 + 4G @ 2 w0 @ 2 w00 dx dy @x2 @y2 @y2 @x2 @x@y @x@y

ZZ

w0 w00 2 V2 :
(5.3.25)

In accordance with results of Section 4.2 (see Conclusion in subsec. 4.2.5), one connects with the lost of stability of the plate the least positive eigenvalue of the following problem ( hi whi ) 2 R V2 whi 6= 0 a(2) (whi w) + hi bqh (whi w) = 0 w 2 V2 h (5.3.26)

bqh being the bilinear symmetric form on V2 V2 determined by the formula bqh (w0 w00 ) =
0 00 0 00 N1 (qh ) @w @w + N2 (qh ) @w @w @x @x @y @y 00 0 00 0 + N3 (qh ) @w @w + @w @w w0 w00 2 V2 : (5.3.27) @x @y @y @x dx dy

ZZ

Taking into account the Holder inequality, we have

jbqh (w0 w00 )j 2

W41 ( ). Moreover, (5.3.28) and Lemma 5.3.1 imply the following statement.

w0 w00 2 V1 : i=1 (5:3:28) ; 1 Thus, bqh 2 L2 W4 ( ) W41 ( ) R , i.e., bqh is a continuous form on W41 ( )
)

3 X

kNi (qh ) kL2 ( ) kw0 kW41 ( ) kw00 kW41 (

Lemma 5.3.2 Let the hypotheses of Lemma 5.3.1 be satis ed and let a bilinear form bqh be de ned by the formula (5.3.27). Then, for p > 2, the function h !

bqh ;is a continuous mapping of Yp equipped with the Wp1 ( )-weak topology into L2 W41 ( ) W41 ( ) R :
Since the bilinear form a(2) is coercive in V2 (see Theorem 4.1.2), zero cannot h be an eigenvalue of the problem (5.3.26), so that we can consider the following problem, instead of (5.3.26), ( hi whi ) 2 R V2 whi 6= 0 (2) w 2 V2 : hi ah (whi w) + bqh (whi w) = 0 (5.3.29)

5.3. Optimization problems for plates (control by the function of the thickness) 313

If hi 6= 0, then

hi =

hi being an eigenvalue of the problem (5.3.26). In this connection, we stress that, for some loads Q, one can get hi = 0, i.e., zero is an eigenvalue of the problem (5.3.29). By virtue of Theorem 1.6.2, the imbedding of V2 into W41 ( ) is compact. So, Theorem 1.5.8 yields that the problem (5.3.29) has a countable number of eigenvalues j h1 j j h2 j lim = 0: (5:3:31) i!1 hi

hi

(5:3:30)

Thus, we can de ne the function h ! (h) = f hi g1 which maps an i=1 element h 2 Yp into h 2 `1 0 . We remind that `1 0 is the normed space of bounded number sequences converging to zero and

k h k`1 0 = sup j hi j : i

(5:3:32)

Analogously to the proof of Lemma 2.5.1, we deduce h ! a(2) to be a conh tinuous mapping of Yp equipped with the Wp1 ( )-weak topology into L2 (V2 V2 R) for p > 2. Combining this with Lemma 5.3.2 and making use of Theorem 2.5.2, we get the following statement.

Lemma 5.3.3 Let the hypotheses of Lemma 5.3.1 be satis ed and let bilinear

forms a(2) and bqh be de ned by the formulas (5.3.25) and (5.3.27). Then, the h function h ! h is a continuous mapping of Yp endowed with the Wp1 ( )-weak topology into `1 0 .

Optimization problem

Suppose we are given functionals k (h ) such that (h ) ! k (h ) is a continuous mapping of Yp `1 0 9 > equipped with the topology generated by the product of the= Wp1 ( )-weak topology and `1 0 -strong one into R, p > 2,> k = 0 1 2 : : : m. De ne a set of admissible controls as Uad = h j h 2 Wp1 ( ) khkWp1 ( ) c h h ^ h (h h ) 0 k = 1 2 : : : m : k Here, ^ p > 2, c, h, and h are positive numbers such that e1 < h < ^ h < e2 , e1 and e2 are the positive constants from (2.1.2).

(5:3:33)

(5.3.34) (5:3:35)

314

Chapter 5. Optimization of deformable solids

Since the sequence fhn g is bounded in Wp1 ( ), p > 2, we can choose a subsequence fhmg such that ~ hm ! h weakly in Wp1 ( ) (5.3.39) hm ! ~ strongly in C ( ): h (5.3.40) (5.3.39) and Lemma 5.3.3 yield that in `1 0: (5:3:41) ~ hm ! h Further, making use of (5.3.33) and (5.3.37){(5.3.41), one can easily see that the ~ function h0 = h is a solution of the problem (5.3.36). An approximate solution of the problem (5.3.36) can be constructed by employing results of subsecs. 2.7.2 and 2.7.3. Consider some realizations of the functionals k . Introduce the notation + = sup : (5:3:42) hi h Assume that h h is the largest positive eigenvalue of the problem (5.3.29) (see Theorem 1.5.5). The critical load under which the plate loses stability equals Q (see Fig. 5.3.3). We point out that, if + = 0, then the plate + h h does not lose stability. More precisely, in this case the plate loses stability if ;Q is substituted for Q, i.e., the load has the opposite direction. Suppose we need that the plate loses stability under a load that is not less than a given one. Then, the functional 1 can be taken in the following form (5.3.43) (h ) = c ; 1
1 + > 0. Then +

Theorem 5.3.1 Let the hypotheses of Lemma 5.3.3 be satis ed and let (5.3.33) hold. Assume that a nonempty set Uad is de ned by the relations (5.3.34) and (5.3.35). Then, the problem (5.3.36) has a solution. Proof. Let fhng be a minimizing sequence, i.e., fhn g Uad (5.3.37) lim 0 (hn hn ) = h2U 0 (h h ) : inf (5.3.38) n!1
ad

The optimization problem consists in nding a function h0 such that h0 2 Uad inf 0 (h h ) : (5:3:36) 0 (h0 h0 ) = h2U
ad

formulas

c being a positive number. Of course, the plate must not be destroyed under the load cQ. Streses in the midplane of the plate under the load cQ are given by the
11 h

( = cN1h qh )

22 h =

cN2 (qh ) h

12 h

( = cN3h qh ) :

(5:3:44)

5.3. Optimization problems for plates (control by the function of the thickness) 315

Here, Ni (qh ) are determined by the relations (5.3.21), and qh is a solution of the problem (5.3.20). ( Let ij )h be the averaging of the stresses ij h in the x and y coordinates, being the radius of the averaging kernel. In view of (5.1.6), the functional 2 can be taken in the form
2 (h h ) = (x y)2

max
2 X

2 X

i j =1

() ij ij h (x y ) ( ) (x

i j m n=1

ijmn ij h

y) mn h (x y) ; 1:

()

(5.3.45)

Then, the condition 2 (h h ) 0 means that the plate whose thickness is de ned by the function h will not be destroyed under the load cQ. Note that in some cases the restriction on the strength is not introduced in the formulation of the problem. Then, one has to verify whether the solution obtained satis es this restriction. Now, provided ZZ (h h ) = h dx dy (5:3:46) 0 the problem (5.3.36) means that one searches for a plate of minimal weight which loses stability under a load that is not less than a given one, and meets the restriction on strength. If we set
+ 0 (h h ) = h () 1 (h h ) = 2

ZZ

h dx dy ; c1

(5:3:47)

in which ij h is generated by the load Q , then the problem (5.3.36) corresponds + h to the maximization of the load that causes the plate to lose stability under the restrictions on the weight and strength.

c1 being a positive constant, and assume

to be de ned by the formula (5.3.45)

In accordance with results of subsec. 4.1.5, the determining of the frequencies of free oscillations reduces to the following problem ( ih uih ) 2 R V ah (uih v) = bh (uih v) v2V (5.3.48) ih where the bilinear forms ah and bh are de ned by the formulas (4.1.24) and (4.1.53), and ih = 1i (see (4.1.55)). By virtue of Theorem 4.1.3, to every function h 2 Yp there corresponds the element h = ( 1h 2h : : : ) 2 `1 0 .

5.3.3 Optimization of frequencies of free oscillations

316

Chapter 5. Optimization of deformable solids

Lemma 5.3.4 Let the hypotheses of Theorem 4.1.3 be ful lled. Then, for p > 2, the function h ! h = ( 1h 2h : : : ) is a continuous mapping of Yp endowed with the Wp1 ( )-weak topology into `1 0 : Proof. Similarly to the proof of Lemma 2.5.1, one nds that, for p > 2, the function h ! ah is a continuous mapping of Yp endowed with the Wp1 ( )-weak topology into L2 (V V R), and the function h ! bh is a continuous mapping of Yp endowed with the Wp1 ( )-weak topology into L2 (W21 ( ) W21 ( ) R). Now, Lemma
5.3.4 is a consequence of Theorem 2.5.2. The problem of optimization of frequencies of free oscillations reduces to the problem (5.3.36) in which h is a solution of the problem (5.3.48). By using Lemma 5.3.4, analogously to the proof of Theorem 5.3.1, one gets the following statement.

Theorem 5.3.2 Let the hypotheses of Theorem 4.1.3 be ful lled and the functionals k , k = 0 1 : : : m, be given and satisfy the condition (5.3.33). Suppose that a nonempty set Uad is determined by (5.3.34) and (5.3.35), h being de ned by the solution of the problem (5.3.48). Then, the problem (5.3.36) has a solution.

We point out that the problem of optimization of frequencies of free oscillations can correspond to the minimization of the weight of the plate under restrictions on the spectrum, as well as to the maximization of the rst natural frequency 1h (i.e., the minimization of 1h = 1h ) under restrictions on the 1 weight and spectrum. Restrictions on the spectrum can mean that the frequencies of free oscillations of the plate do not get into some intervals, which are determined by the dynamic load. Some examples of construction of the functionals k were presented above, in subsecs. 5.3.2 and 2.6.1.

Assume that, in di erent periods of time, a plate is subject to longitudinal and transversal forces as well as certain periodical (in time) forces. In view of this, the plate should satisfy a system of restrictions on the strength, stability, and frequency of free oscillations. The goal functional can correspond to the weight of the plate, which is to be minimized. The corresponding optimization problem reduces to a combined problem (see Section 2.9). Combined problems for beams and plates were examined by Seyranian (1973, 1976, 1977). Using results of subsecs. 5.3.1{ 5.3.3 and Section 2.9, one can derive the existence of a solution of a combined optimization problem for plates and the convergence of the approximate solutions of this problem which are constructed by using the technique of subsec. 2.8.2. Taking regard of results of Section 2.10, one can optimize plates for a class of loads. Notice that problems of optimization of beams and plates for a class of loads were studied by Banichuk (1980).

5.3.4 Combined optimization problem and optimization for a class of loads

5.4. Optimization problems for shells

:::

317

5.4 Optimization problems for shells (control by the functions of the midsurface and of the thickness)
\I rode over to see her once every week for a while and then I gured it out that if I doubled the number of trips I would see her twice as often" | O. Henry \The Pimienta Pancakes" For optimization of shells, one can take as controls the function of the thickness of the shell and the function determining the midsurface (see (4.5.1)) for shallow shells, the control can be the shape of the plan of the shell, i.e., the shape of the projection of the midsurface of the shell on the (x y) plane (see Fig. 4.7.1). Naturally, other kinds of controls (e.g., the optimal sti ening of the shell) are also applicable. However, in this section we study control by the shape of the midsurface and the thickness only, while in Section 5.5 we investigate the problem of control by the shape of the plan of a shallow shell.

5.4.1 Problem of optimization of a shell of revolution with respect to the strength

According to results of Sections 4.6 and 4.8, the problem on the stress-strain state of a shell of revolution reduces to nding a function of displacements ! = (~ v w) ~ u~ ~ such that !2V ~ ah (~ !) = (f !) ! !2V (5:4:1) f being an element from V determined by the load acting on the shell, and the bilinear form ah being de ned by the formula (4.6.13). The midsurface of the shell of revolution is determined by a function r(z ) which is called the meridian function (see Fig. 4.6.1). To study the optimization problem, we need the property of continuous dependence of a solution of the problem (5.4.1) of the function of the thickness of the shell h(' z ), of the meridian function r(z ), and of the load f . Therefore, let us proceed study this property. De ne sets ^ G1 = h j h 2 C ( ) h h h hj'=0 = hj'=2 (5.4.2) 3 ( 0 L]) r r r G2 = r j r 2 C ^ (5.4.3) where ^ ^ ^ h h r r are positive numbers h < h r < r ^ (5:4:4)

Direct problem and continuous dependence

318

Chapter 5. Optimization of deformable solids

and is de ned by (4.6.1). The sets G1 and G2 are equipped with the topologies generated by the topologies of the spaces C ( ) and C 3 ( 0 L]), respectively. In view of (4.6.3){(4.6.5), (4.6.10), and (4.6.13), the bilinear form ah depends not only on h, but also on the function r, so that we will denote it as ahr , i.e., ah = ahr : (5:4:5) The solution of the problem (5.4.1) depends obviously on h, r, and the load f . Hence, by (5.4.5), the problem (5.4.1) can be rewritten as follows: !hrf 2 V ahr (!hrf !) = (f !) ! 2 V: (5:4:6) Lemma 5.4.1 Let the conditions (4.3.6){(4.3.8) be satis ed, let the space V be endowed with the norm (4.6.12) and satisfy (4.6.11). Suppose that sets G1 and G2 are de ned by the formulas (5.4.2){(5.4.4). Let also (5.4.5) hold and let a bilinear form ah be de ned by the relations (4.5.10), (4.6.3){(4.6.5), (4.6.10), and (4.6.13). Then, for any (h r f ) 2 G1 G2 V , the problem (5.4.6) has a unique solution !hrf , and the function (h r f ) ! !hrf de ned by this solution is a continuous mapping of G1 G2 V into V: Proof. Theorem 4.6.1 implies that, for all (h r) 2 G1 G2 , the bilinear form ahr is symmetric, continuous, and coercive. Thus, by virtue of the Riesz theorem, for any (h r f ) 2 G1 G2 V , there exists a unique solution to the problem (5.4.6). To prove the continuity of the function (h r f ) ! !hrf , we use Theorem 2.11.1. The operators Pi from (4.6.13) determined by the formulas (4.6.3){(4.6.5) and (4.6.10) depend obviously on r, i.e., we can write Pi = Pi(r). It is easy to verify that r ! Pi(r) is a continuous mapping of G2 into L(W L2 ( )), the space W being de ned by (4.6.8). The formulas (4.5.10) and (4.6.13) yield that cij and Dlm depend on h, i.e., (h (h cij = c(h) and Dlm = Dlm) . Obviously, h ! c(h) and h ! Dlm) are continuous ij ij mappings of G1 into C ( ). Now, Theorem 2.11.1 yields (h r f ) ! !hrf to be a continuous mapping of G1 G2 V into V . (In this case, we set B1 = C ( ) C 3 ( 0 L]) and Q1 = G1 G2 .) We will control by the function of the thickness of a shell of revolution h(' z ) and by the shape of its midsurface, i.e., the meridian function r(z ). To every pair of functions (h r) 2 G1 G2 we place in correspondence a load fhr acting on the shell, i.e., we suppose that we are given a function (h r) ! fhr such that (h r) ! fhr is a continuous mapping of G1 G2 into V : (5:4:7) Now, by (5.4.6), the function of displacements of the midsurface of the shell !hr = !hrfhr is determined as the solution of the problem !hr 2 V ahr (!hr !) = (fhr !) ! 2 V: (5:4:8)

Optimization problem

5.4. Optimization problems for shells

:::

319

Further, suppose we are given functionals k (h r !) such that (h r !) ! k (h r !) is a continuous mapping of G1 G2 V into R, k = 0 1 2 : : : m. De ne a set of admissible controls as

(5:4:9)

^ Uad = (h r) j h 2 Wp1 ( ) khkWp1 ( ) ch h h h h '=0 = h '=2 r 2 W24 ((0 L)) krkW24 ((0 L)) cr r r r k (h r !hr ) 0 k = 1 2 : : : m : (5.4.10) ^ Here, !hr is the solution of the problem (5.4.8) and 9 ^ p > 2, ch and cr are positive numbers, h, h, r, and r are the= ^ (5:4:11) constants from (5.4.2) and (5.4.3) satisfying the conditions (5.4.4). The functionals k in (5.4.10) can be considered as restrictions on the strength and sti ness of the shell (see subsec. 5.1.2). The optimization problem consists in nding a pair (h0 r0 ) such that (h0 r0 ) 2 Uad
0 (h0 r0

lations (5.4.7), (5.4.9) hold. Assume that a nonempty set Uad is de ned by the formulas (5.4.10) and (5.4.11). Then, the problem (5.4.12) has a solution.

Theorem 5.4.1 Let the hypotheses of Lemma 5.4.1 be ful lled and let the reProof. Let fhn rng1=1 be a minimizing sequence, i.e., n fhn rn g1=1 Uad n

!h0r0 ) = (h rinfU )2

ad

0 (h

r !hr ) :

(5:4:12)

lim (h r ! ) = inf 0 (h n!1 0 n n hn rn (h r)2Uad

r !hr ) :

(5.4.13) (5.4.14)

Since the set Uad is bounded in Wp1 ( ) W24 ((0 L)), the imbedding theorem implies that one can choose a subsequence fhm rm g1=1 of the sequence m fhn rn g1=1 such that n ~ ~ hm ! h weakly in Wp1 ( ) hm ! h strongly in C ( ) rm ! r weakly in W24 ((0 L)) ~ rm ! r strongly in C 3 ( 0 L]): (5.4.15) ~ Obviously, ~ 2 G1 and r 2 G2 . Now, making use of Lemma 5.4.1, (5.4.7), h ~ ~ (5.4.9), and (5.4.13){(5.4.15), it is easy to see that the pair (h0 = h r0 = r) is a ~ solution of the problem (5.4.12). The theorem is proven. Approximate solutions of the problem (5.4.12) can be constructed by using the technique of subsec. 2.11.3. Just as above, due to results of Section 2.11, one

320

Chapter 5. Optimization of deformable solids

can investigate and solve eigenvalue optimization problems which are connected with the stability and oscillations of the shell, as well as combined problems in the case when the control is the function of the thickness of the shell and the shape of the midsurface. Optimization problems on the strength and rigidity of cylindrical shells of variable thickness that are subject to an axially symmetric load were studied by Medvedev and Totskii (1984b). Now, let us consider an optimization problem on the stability of a cylindrical shell.

5.4.2 Optimization according to the stability of a cylindrical shell subject to a hydrostatic compressive load
Direct problem
Let a cylindrical shell be subject to a load that is uniformly distributed on the surface with intensity ; . According to results of subsec. 4.10.2, the problem of stability of the shell reduces to determining the least positive eigenvalue of the following problem 2 R ! 2 V ! 6= 0 ~ ~ ah (~ !) = bh ! (~ !) ! 2 V: ! (5:4:16) ^ ! Here, the form ah is de ned by the formulas (4.5.10), (4.6.3){(4.6.5), (4.6.10), and (4.6.13) with r = const = R, where R is the radius of the shell. If the thickness of the shell is small as compared to its radius, the bilinear form bh ! can be de ned in ^ the form (4.10.15), and the forces T1 , T2 , and S can be found from the membrane theory of shells. From this theory it follows that, in the case under investigation, T1 = S = 0 and T2 = ;R. Then, the form bh ! takes the form ^

bh ! (!0 !00 ) = ^

Z LZ 2
0 0

where !0 = (u0 v0 w0 ) and !00 = (u00 v00 w00 ).

@w0 @w00 dz d' @' @'

(5:4:17)

We will treat the function of the thickness of the shell h as a control and suppose h to depend only on z . Therefore, we take a set of admissible controls in the form

Optimization problem
n

Uad = h j h 2 W21 ((0 L)) khkW21 ((0 L)) ch

^ h h h

ZL
0

h dz c f (h) 0 : (5.4.18)

Here, f is the function of restrictions on the strength. Let 1h be the least positive eigenvalue of the problem (5.4.16), and let the goal functional f1 have the form f1 (h) = ; 1h : (5:4:19)

5.4. Optimization problems for shells

:::

321

The optimization problem consists in determining a function h0 such that

h0 2 Uad

f1(h0 ) = h2U f1 (h): inf

ad

(5:4:20)

Since the imbedding of W21 ((0 L)) into C ( 0 L]) is compact, if the set Uad is nonempty, then the problem (5.4.20) has a solution. In the work by Medvedev (1980), the problem of optimization of an orthotropic cylindrical shell clamped at the edges and subject to a hydrostatic compressive load was numerically solved. The parameters of the shell were L = 3 m, R = 0:75 m, 1 = 2 = 0:11, E1 = E2 = 3:7 1010 Pa, and G = 0:75 1010 Pa. To construct an approximate solution of the direct problem, i.e., to calculate 1h , the spaces Vn were constructed as tensor products of splines and trigonometric polynomials (see subsec. 5.2.3). In solution of the nite-dimensional optimization problem, the function of the thickness of the shell z ! h(z ) was searched for in the class of piecewise-a ne functions which are continuous on 0 L] and symmetric with respect to the point z = L . More precisely, optimal forms were searched 2 for in the class of the functions pictured on Fig. 5.4.1 by a continuous line. The point line corresponds to the shell of a constant thickness. The values q1 and q2 shown in Fig. 5.4.1 were used as optimization parameters q3 was chosen so that the weights (volumes) of the shells of constant and variable thickness are equal, the angle being considered to be xed (see Fig. 5.4.1).
h 102 / (2 R )
q2 q1 q3
a

0.5

0.2

0.4

0.6

0.8

z/ L

Figure 5.4.1: Function of thickness of a cylindrical shell q1 and q2 are parameters of optimization Let q = (q1 q2 ). Obviously, in the case under investigation, the eigenvalues of the problem (5.4.16) depend on q. Hence, we denote them as i (q) and order them in such a way that 0 < 1 (q)
2 (q )

(5:4:21)

The problem of maximization of the function q ! 1 (q) on the set Uad h of the form Uad h = q = (q1 q2 ) j q 2 R2 q2 q1 h (5:4:22)

322

Chapter 5. Optimization of deformable solids

was solved. (The set of admissible controls Uad depends here on h, so that we ^ denoted it as Uad h .) The upper restriction on the thickness h is not included in Uad h , since ^ is supposed to be chosen su ciently large, so that the optimal h solution always satis es it. We denote by qh a solution of the problem

qh 2 Uad h

1 (qh ) =

q2Uad h

sup

1 (q ):

(5:4:23)

Fig. 5.4.2 show the dependence h ! i (qh ), where i (qh ) are the eigenvalues appropriate to qh (see (5.4.21)). Lines 1{5 on Fig. 5.4.2 determine the values 1 (qh ), : : : , 5 (qh ). The value h = 10;2 m corresponds to a shell of a constant thickness. For this shell, as Fig. 5.4.2 shows, the rst eigenvalue is of simple multiplicity. The decrease of h causes the increase of 1 (qh ) and the multiplicity of this eigenvalue.
l 0.75 102 , N / m

2.6 2.4 2.2 2.0 18 . 16 .

5 4

3 2 1
05 . 0.6 0.7

0.4

0.8

0.9

h 102 , m

Figure 5.4.2: Eigenvalues of the optimal shells as functions of the parameter h Thus, for example, for h = 0:75 10;2 m the multiplicity of 1 (qh ) equals 5, and for the smaller values of h the value of 1 (qh ) is almost the same and its multiplicity does not change. We point out that the results obtained are in agreement with the engineering approach to optimization of constructions, which demands the optimal construction to have \general" and \local" forms of loss of stability the latter corresponds to the requirement that the multiplicity of the rst eigenvalue must be greater than one. Medvedev and Totskii (1984a) examined the problem of maximization of the rst eigenvalue of the problem (5.4.16) on the set

Uad = h j h 2 Sp(1) (

38

0 L])

5.5. Control by the shape of a hole

:::

323

o ^ h L ; y = h L + y for y 2 0 L=2] h h h (5.4.24) 2 2

provided bh ! is de ned by (5.4.17). Here, Sp(1) ( 38 0 L]) is the set of piecewise^ a ne functions which correspond to the partition of the interval 0 L] into 38 parts (see subsec. 5.2.5), and the partition was taken uniform. As (5.4.24) shows, the optimal solution was searched for in the class of the functions symmetric with respect to the point z = L , and the optimization was carried out by twenty 2 parameters. The calculation was done for the orthotropic cylindrical shell clamped on the edges in the case when L = 3 m, R = 0:75 m, 1 = 2 = 0:11, E1 = E2 = ^ 3:7 1010 N=m2 , G = 0:75 1010 N=m2 , h = 0:02 m, and h = 0:08 m. On Fig. 5.4.3, shown is the function h0 (z ) obtained, which maximizes the functional h ! 1h , i.e., the rst eigenvalue (the critical load) on the set Uad from (5.4.24). (More precisely, h0 is a local maximum since the function h ! 1h is not ~ convex.) As compared to the shell of the constant thickness h(z ) = 0:04 m for all z 2 0 L] and of the same weight as one of the shell with the thickness h0 (z ), the critical load increases approximately twice, i.e., 11hh0 2. For the optimal shell ~ with the thickness h0 , the multiplicity of the rst eigenvalue 1h0 equals twelve.
h 0 102 / (2 R )
4 2
0

0.2

0.4

0.6

0.8

z/ L

Figure 5.4.3: Function of thickness of the optimal cylindrical shell

5.5 Control by the shape of a hole and by the function of thickness for a shallow shell
5.5.1 Problem of optimization according to the strength
Consider the optimization problem for a shallow shell which is described by the shear model (see subsec. 4.11.2). Let spaces M1 and N1 be de ned by the relations (2.12.41) with l = 1, and equipped with topologies as in subsec. 2.12.4. Then the condition (2.12.10) is satis ed. To each element q 2 M1 we place in correspondence a two-connected domain q whose inner boundary S1 is de ned in polar coordinates by the function ! q( ), and whose external boundary S2 is supposed to be xed (Fig. 5.5.1) and determined by a function ! q1 ( ) such that

324

Chapter 5. Optimization of deformable solids

q1 2 C 1 ( 0 2 ]), q1 (0) = q1 (2 ), and dq1 (0) = dq1 (2 ) d d r2 de ned in (2.12.41).

y2

202

min ] q1 ( ) > r2

S1
q(a )

q1 (a )

S2

y1

Figure 5.5.1: Two-connected domain q By using the mapping Pq de ned in subsec. 2.12.4, one de nes a di eomorphism of the C 1 class of the set q onto , where is determined by (2.12.36). Further, let

G = h j h 2 Wp1 ( ) e1 h e2 khkWp1 ( ) e3 (5:5:1) where e1 , e2 , and e3 are positive numbers and p > 2. The set G is endowed with the topology generated by the Wp1 ( )-weak one. Then, the condition (2.12.4) holds. ; ; Let V ( q ) and V ( ) be closed subspaces of W21 ( q ) 5 and W21 ( ) 5 , respectively, which satisfy the condition (4.11.8) and the following condition of concordance: the mapping u ! u Pq is an isomorphism of V ( ) onto V ( q ). (5:5:2)

Suppose we are given functions k1 and k2 on such that k1 2 L1 ( ), k2 2 L1( ). Now, to every pair (h q) 2 G M1 we place in correspondence the bilinear form ahq which is generated by the strain energy of the shell. This form is de ned on V ( q ) V ( q ) and determined by the right hand side of (4.11.9), ; provided h Pq is substituted for h, q is substituted for , and R1 1 = k1 Pq , ;1 R2 = k2 Pq . De ne a bilinear form Pq ahq on V ( ) V ( ) by the relation Pq ahq (u v) = ahq (u Pq v Pq ) u v 2 V ( ): (5:5:3)

5.5. Control by the shape of a hole

:::

325

i.e., the condition (2.12.6) is ful lled. Assume that we are given a continuous mapping G M1 3 (h q) ! B (h q) 2 (V ( )) . From the physical point of view, for each control, i.e., for each shape of the hole which is de ned by the curve S1 (see Fig. 5.5.1), and for every function of thickness h Pq , the mapping B de nes the load B (h q) acting on the shell. (More precisely, B (h q) is the image of the load corresponding to the change of variables de ned by the mapping Pq .) In particular, B can be a constant mapping, which corresponds to the case when the load, as an element of the space (V ( )) , does not depend on control. Given a control (h q) 2 G M1 , determining the stress-strain state of the shell reduces to nding a function uhq such that ; uhq 2 V ( q ) ahq (uhq v) = B (h q) v Pq;1 v 2 V ( q ): (5:5:5) The problem (5.5.5) is equivalent to the following one: Find a function uhq such ~ that uhq 2 V ( ) ~ Pq ahq (~hq w) = (B (h q) w) w 2 V ( ) u (5:5:6) in this case uhq = uhq Pq;1 : ~ (5:5:7) De ne a set of admissible controls as Uad = (h q) j h 2 G q 2 N1 k (h q uhq ) 0 k = 1 2 : : : l : (5:5:8) ~ Here, (h q u) ! k (h q u) are continuous mappings of G N1 V ( ) into R. The functionals k can play the role of restrictions on the strength as well as those on the hole in the shell (for example, one can require that the area of the hole be not less than a given one, or that the hole contain a certain subdomain, etc.). The set Uad is supposed to be nonempty, and the goal functional is assumed to be of the form Z Z ; J (h q) = (h Pq )(y) dy = h(x) det Pq;1 0 (x) dx: (5:5:9) Here, Pq is the derivative (the Jacobi matrix) of the mapping Pq;1 . The optimization problem of nding a pair (h0 q0 ) such that (h0 q0 ) 2 Uad J (h0 q0 ) = inf J (h q) (5:5:10) means that one searches for a law of the change of the thickness of the shell and a shape of the hole for which the weight of the shell is minimal provided restrictions on the strength and other requirements are ful lled. In view of Theorem 2.12.1, under the above assumptions the problem (5.5.10) has a solution.
(h q)2Uad

Pq ahq and easily concludes that (h q) ! Pq ahq is a continuous mapping of G M1 into L2 (V ( ) V ( ) R),

By using the rule of change of variables, one obtains the explicit form of (5:5:4)

;1 0

326

Chapter 5. Optimization of deformable solids

Let fHn g be a sequence of nite-dimensional subspaces of Wp1 ( ), let fVn g be a ~ sequence of nite-dimensional subspaces of W22 (0 2 ), and let the limit density conditions lim inf kh ; z kWp1 ( ) = 0 h 2 Wp1 ( ) (5.5.11) n!1 z2H
n!1 '2Vn

5.5.2 Approximate solution of the optimization and direct problems

lim inf kq ; 'kW22 (0 2 ) = 0 f

f q 2 W22 (0 2 )

(5.5.12)

be ful lled. f We recall that W22 (0 2 ) is the subspace of periodic functions of the space 2 (0 2 ) (see subsec. 2.12.4). The spaces Vn meeting the condition (5.5.12) can W2 be constructed by trigonometric polynomials or cubic periodical;splines. Write Qn = Hn Vn and de ne an approximate solution h(n) q(n) of the problem (5.5.10) in the form

h(n) q(n) 2 Qn \ Uad

J h(n) q(n) = (h q)2inf \U J (h q): Qn ad

(5:5:13)

For studying the solvability of the problem (5.5.13) and the convergence of ; the approximate solutions h(n) q(n) to the explicit solution (h0 q0 ), one can use results stated above (see, e.g., Section 2.3). The optimization problem (5.5.13) can be solved by the iteration method. To this end, for every iteration one has to solve the direct problem (5.5.5) in a two-connected domain q of a rather complicated shape, the domain changing at each iteration. Let us show that an approximate solution of the problem (5.5.5) can be obtained by the nite element method, by replacing the domain q with a certain two-connected domain the connected components of the boundary of which are boundaries of polygons. De ne a mapping n ! k(n) of the set of natural numbers N into itself such that k(n) ! 1 as n ! 1. To each n 2 N we place in correspondence two closed, piecewise-straight lines S1n and S2n which are the boundaries of the polygons ) ) ) ) with vertices A(in , A(in , i = 1 2 : : : k(n), having the polar coordinates ( (in (in ), 1 2 1 1 (i) (i) ) determined by the formulas (see Fig. 5.5.2) ( 2n 2n (i) = 2 i (i) = q 2 i 1n k (n) 1n k(n) 2 i 2 i (i) = q (i) = i = 1 2 : : : k(n): (5.5.14) 1 k (n) 2n 2n k (n) If 1n ( ) and 2n ( ) are the functions determining the boundaries S1n and S2n , then well-known results on the estimation of the remaining term of an interpolation spline of the rst power (see Zavialov et al. (1980)) imply that 1 1 (5:5:15) 1n ! q in W1 ((0 2 )) 2n ! q1 in W1 ((0 2 )):

5.5. Control by the shape of a hole

y2
( A21) n ( A2 2 ) n 1 A1(n )

:::

327

(2) 1n

2n

a( )

S2 y1

r 1na ( )

q1

S1n S1

S2 n

Figure 5.5.2: Domain q and polygons Denote by n the two-connected domain bounded by the lines S1n and S2n . By using the functions n = ( 1n 2n ) and the technique of subsec. 2.12.4, one 1 constructs a di eomorphism Pn = P n of the W1 class of the set n onto , i.e.,
1 ; 1 Pn 2 W1 ( n ) Pn 1 2 W1 ( n ): (5:5:16) ; By virtue of (5.5.16), the mappings Pn and Pn 1 are Lipschitz continuous ones, so that the mapping u ! u Pn is an isomorphism of W21 ( ) onto W21 ( n ) (see Necas (1967)), and because of (5.5.2) u ! u Pn is an isomorphism of V ( ) onto V ( n ). Now, to every n we place in correspondence the bilinear, symmetric, continuous, and coercive form an on V ( n ) V ( n ) determined by the relation (4.11.9) ; provided h Pn is substituted for h, n is substituted for , and R1 1 = k1 Pn , ;1 R2 = k2 Pn . Let also Pn an be the image of the form an under the mapping Pn ,

given by the formula

Pn an (u v) = an (u Pn v Pn )
By using (5.5.15), one can verify that

u v 2 V ( ):

(5:5:17) (5:5:18)

Pn an ! Pq ahq u(n) 2 V ( n )

in L2 (V ( ) V ( ) R) as n ! 1:
;

Consider the following problem: Find a function u(n) such that

; an u(n) w = B (h q) w Pn 1

w 2 V ( n ): (5:5:19)

328

Chapter 5. Optimization of deformable solids

By virtue of (5.5.17), for the function u(n) solving the problem (5.5.19), the following relation holds
; Pn an u(n) Pn 1 v = (B (h q) v)

v 2 V ( ):

(5:5:20)

(5.5.6), (5.5.7), (5.5.18), (5.5.20), and Theorem 1.8.1 yield that ; u(n) Pn 1 ! uhq = uhq Pq;1 in V ( ) as n ! 1: ~ Since the boundary of n consists of piecewise-straight lines, by applying the nite element method one can construct approximate solutions of the problem (5.5.19) which converge in V ( n ) to the explicit solution u(n) of this problem. For the nite shear model of a shallow shell, the frequencies of natural oscillations are determined as solutions of the eigenvalue problem (4.11.11). According to results of subsec. 5.5.1, to every pair (h q) 2 G M1 there correspond a domain q and bilinear forms ahq and bhq de ned by the right hand sides of the relations (4.11.9) and (4.11.12), respectively, provided h Pq is substituted for h, q is ; ; substituted for , and R1 1 = k1 Pq , R2 1 = k2 Pq . By Pq bhq we will denote the bilinear, symmetric, continuous form on (L2 ( ))5 (L2 ( ))5 which is the image of the form bhq under the mapping Pq : Pq bhq (u v) = bhq (u Pq v Pq ) u v 2 V ( ): (5:5:21) It is easy to verify that (h q) ! Pq bhq is a continuous mapping of G M1 into (5:5:22) L2 ((L2 ( ))5 (L2 ( ))5 R). Now, consider the following problem: Given a pair (h q) 2 G M1 , nd eigenvalues (i) of the problem hq (5.5.23) Results of subsec. 4.11.2 imply the existence of a sequence of eigenvalues of the problem (5.5.23) such that f (i) g1 = hq 2 `1 0, (i) > 0 for any i, and hq i=1 hq
(1)

5.5.3 Problem of optimization of eigenvalues

u(i) 2 V ( q ) u(i) 6= 0 hq hq (i) a (u(i) v ) = b (u(i) v ) hq hq hq hq hq

(2) (3)

(i)

hq

2R v 2 V ( q ):

hq

hq

hq

Further, suppose we are given continuous mappings `1 0 3 ! Ai 2 R, i = 1 2 : : : k, and the set of admissible controls Uad = (h q) j (h q) 2 G N1 Ai hq 0 i = 1 2 : : : k : (5:5:24)

lim (i) = 0: i!1 hq

5.5. Control by the shape of a hole

:::

329

We assume that Uad is nonempty and the goal functional : (h q) ! (h q) is a continuous mapping of G N1 into R. Taking into account (5.5.4) and (5.5.22), from Theorem 2.12.2 we deduce that there exists a pair (h0 q0 ) such that (h0 q0 ) 2 Uad (h0 q0 ) = inf (h q): (5:5:25)
(h q)2Uad

5.5.4 Approximate solution of the eigenvalue problem

Let us show that to get approximate eigenvalues of the problem (5.5.23), one can replace the domain q by one with a boundary consisting of piecewise-straight lines. Let f n g be a sequence of domains which are constructed as in subsec. 5.5.2, let fPn g be a sequence of di eomorphisms meeting the condition (5.5.16), and let fan g be the corresponding sequence of bilinear forms on V ( n ) V ( n ) de ned in subsec. 5.5.2. Further, for each n, we de ne a bilinear continuous form bn on (L2 ( n ))5 (L2( n ))5 by the right hand side of the relation (4.11.12), with h Pn substituted for h and n substituted for . Denote by Pn bn the image of the form bn under the mapping Pn : Pn bn (u v) = bn (u Pn v Pn ) u v 2 (L2 ( ))5 : It is easy to see that Pn bn ! Pq bhq in L2 ((L2 ( ))5 (L2 ( ))5 R) as n ! 1: (5:5:26) Now, consider the eigenvalue problem u(i) 2 V ( n ) u(i) 6= 0 (i) 2 R n n (i) a n(u(i) v ) = b (u(i) v ) v 2 V ( n ): (5.5.27) n n n n n The problem (5.5.23) is equivalent to the following one: u(i) 2 V ( ) u(i) 6= 0 (i) 2 R hq (i) P a (i) w = P b (i) w w 2V( ) q hq u hq q hq u where u(i) = u(i) Pq;1 , and problem (5.5.27) is equivalent to the problem hq u(i) 2 V ( ) u(i) 6= 0 (i) 2 R ~n ~n n (i) P a u(i) w = P b u(i) w ~n ~n w 2V( ) n n n n n
; where u(i) = u(i) Pn 1 . (5.5.18), (5.5.26), and Theorem 1.5.9 imply ~n n n o1 n o1 i ! hq = (hq) in `1 0 as n ! 1: n = (i) n

Since the problem (5.5.27) is solved in a domain whose boundary consists of piecewise-straight lines, approximate values of (i) can be calculated by using n the nite element method, by triangulating n .

i=1

i=1

330

Chapter 5. Optimization of deformable solids

5.6 Control by the load for plates and shells

So far we have studied optimization problems for plates and shells in which as a control we took some geometric characteristics (the function of thickness, the equation de ning the form of the midsurface of the shell, the form of the plan of the shell), the load being assumed to be xed. However, if one produces a construction (here and below this word is used for a plate or a shell) having optimal values of geometric characteristics for a xed load, then a change of the load will result in the construction's being no longer optimal. Moreover, in some cases it may happen that this construction will be even \worse than a nonoptimal one." On the other hand, the requirement that the construction be optimal for a wide class of loads which can act on it during the operation, leads to an essential complexi cation of design (computation) of the construction, and may considerably enlarge its weight. Besides, it is impossible to establish a priori all the loads that will act on the construction. So, it seems to be reasonable to use an approach based on the control of the stress-strain state of the construction by control loads. In the latter case, the control loads are adjusted to the basic load and change with the change of it. Static problems of the control by the load, i.e., by control loads, reduce to the control by the right hand sides of elliptic systems. The problems on the optimal heating and cooling of plates and shells reduce to such problems, too (see, e.g., Burak et al. (1984)). Let us consider some problems of such type.

5.6.1 General problem of the control by the load

Setting for the problem
Let a construction be subject to a load f 2 V . The problem on the stress-strain state of the construction reduces to determining a function of displacements !f such that !f 2 V ah (!f !) = (f !) ! 2 V: (5:6:1) Here, the function h is assumed to be xed, while the load f runs over some set. Thus, the solution of the problem (5.6.1) is denoted by !f . Now, suppose that not only the basic load f0, but also a control load f acts on the construction. Then, in accordance with the notations accepted, the problem on the stress-strain state of the construction reduces to determining a function !f0 +f such that

!f0 +f 2 V

ah (!f0 +f !) = (f0 + f !) ! 2 V: !f0 +f = !f0 + !f :

(5:6:2) (5:6:3)

Since ah is a bilinear form, we have

5.6. Control by the load for plates and shells

331

Let L 2 L(V H), where H is a Banach space and z is a given element from H. De ne a goal functional f ! J (f ) as

J (f ) = kL!f0+f ; z kp : (5:6:4) H Here, we set p = 2 if H is a Hilbert space and p 1 otherwise. Suppose U is a space of controls, U is a Hilbert space, and V U V is dense in U , the imbedding of V into U is compact. (5:6:5) Let also Uad be a set of admissible controls such that Uad is bounded and sequentially weakly closed in U: (5:6:6) By identifying the space U with its dual, from (5.6.5) and Theorem 1.5.12 we deduce the imbedding of U into V to be compact. Now, one can easily prove the existence of a function g such that g 2 Uad J (g) = f 2U J (f ): inf (5:6:7)
ad

Depending on the choice of the operator L and the space H (see (5.6.4)), the problem (5.6.7) can correspond to obtaining elds of displacements, stresses, etc., which are nearest to given ones in the mean-square or minimax sense. If in (5.6.4) H is a Hilbert space, p = 2, and the set Uad is convex, then from results of Section 3.3 one can derive necessary and su cient optimality conditions for the problem (5.6.7), as well as utilize the technique of this section for construction of approximate solutions. In this connection, the condition (5.6.5) can be replaced by the following weaker one:

(see Lions (1968), Litvinov (1976)). Let us examine some special problems.

U V is dense in U , the imbedding of V into U is continuous. We stress that, if J is an increasing functional, the set Uad can be unbounded

5.6.2 Optimization problems for plates

Setting of the problem
The bilinear form ah for a plate is de ned by the formula (4.1.24) and V W22 ( ). Let H = L2 ( ) and let L be the operator of imbedding of V into L2 ( ). Then, the problem (5.6.4), (5.6.7) corresponds to searching for a function g 2 Uad such that the de ection function !f0 +g = wf0 +g is the best mean-square approximation of a given function z 2 L2( ). If the operator L acting from V into H = (L2 ( ))3 is de ned by the relation

wf0 +f ! Lwf0 +f = fM1 (wf0 +f ) M2 (wf0 +f ) M3 (wf0 +f )g

332

Chapter 5. Optimization of deformable solids

Mi (wf0 +f ) being the bending moments and torque, determined by formulas (4.1.3) and (4.1.10), then the problem (5.6.4), (5.6.7) corresponds to searching for g 2 Uad
such that the distribution of the bending moments and torque is the best meansquare approximation of a given z = (z1 z2 z3 ) 2 (L2 ( ))3 . Let (x = ai y = bi ) be points of , i = 1 2 : : : n, and let i ,
i = (x y ) j ai ; i

x ai + i bi ; i y bi + i

(5.6.8)

i and i being positive numbers, i and j being disjoint provided i 6= j . Assume that n X f = ; qi i (x y) (5:6:9)

where i is the characteristic function of the set i , i.e.,

i (x y ) =
(

i=1

1 if (x y) 2 i 0 otherwise

(5:6:10)

( )n and ( )H being the scalar products in Rn and H, respectively. By virtue of (5.6.9), the set Uad is mapped onto a set Kad Rn , so that the problem (5.6.7) reduces to determining a vector t such that t 2 Kad I (t) = q2K I (q): inf (5:6:13)
ad

qi are constants, and the vector q = (q1 q2 : : : qn ) is a control. Now, if H is a Hilbert space, p = 2, then, making use of (5.6.3) and (5.6.9), we get the functional (5.6.4) to be a function of vector q which up to a constant summand looks like I (q) = (Aq q)n ; 2(p q)n : (5:6:11) Here, A is an n n matrix with elements aij , p is an n-dimensional vector with components pi , and ; aij = Lw i Lw j H i j = 1 2 ::: n pi = (Lwf0 ; z Lw i )H i = 1 2 ::: n (5.6.12)

Let the set Kad be of the form

corresponds to the lower and upper restrictions on control loads. Assume that the operator L has a left inverse. Then, fLw i gn=1 is a system i of linearly independent functions in H, and the bilinear form q0 q00 ! (Aq0 q00 )n is symmetric, continuous, and coercive in Rn . Now, Theorem 3.1.1 implies that problem (5.6.13), (5.6.14) has a unique solution.

Kad = q j q = (q1 q2 : : : qn ) 2 Rn qi; qi qi+ i = 1 2 : : : n : (5:6:14) Here, qi; and qi+ are given constants such that qi+ > qi; . The relation (5.6.14)

5.6. Control by the load for plates and shells

333

Direct and dual problems

M=

To solve the problem (5.6.13), (5.6.14), we will use duality methods (see, e.g., Sea (1971)). Let

j =(

:::

2n ) 2 R

2n

i
2n X

0 i = 1 2 : : : 2n :
i Ii (q )

(5:6:15) (5:6:16) (5:6:17)

Consider a function : Rn M ! R of the form (q ) = I (q) + where

+ Ii (q) = qi;; qi q q ;

i=1

for i = 1 2 : : : n i;n i;n for i = n + 1 : : : 2n: Taking into account (5.6.11), rewrite (5.6.16) as

(q ) = (Aq q)n ; 2(p q)n + (Bq ; l )2n : (5:6:18) Here, B is a 2n n-matrix and l 2 R2n . The form of the matrix B and the vector l is uniquely determined by the formulas (5.6.16), (5.6.17). One can prove the existence of a saddle point (t ) of the function such that (t ) 2 Rn M min max (q ) = (t ) = max qmin (q ): (5.6.19) q2Rn 2M 2M 2Rn The problem of determining (t ) from the conditions of the left hand side of the equality (5.6.19) is referred to as a direct problem, while the problem of determining (t ) from the condition of the right hand side of (5.6.19) is called a dual problem. It is easy to verify that the vector t from the pair (t ) is a solution of the problem (5.6.13), (5.6.14). Consider the dual problem. Given 2 M , let q be a solution of the problem of determining minq2Rn (q ). The vector q is a solution of the equation Aq = p ; 1 B , i.e., 2 1 q = A;1 p ; 2 B : (5:6:20) Known results on the minimum of a quadratic functional (see Michlin (1970)) imply that (q ) = qmin (q ) = ; (Aq q )n ; (l )2n : 2Rn Now, the dual problem takes the following form: Find a vector such that

2M

;F ( ) = max( ; F ( )) 2M

334 where

Chapter 5. Optimization of deformable solids

or: Find a vector such that

F ( ) = (Aq q )n + (l )2n

(5:6:21) (5:6:22)

2M

F ( ) = min F ( ): 2M

By substituting (5.6.20) into (5.6.21), we conclude that the function F up to a constant summand is determined by the relation ; ;1 1; F ( ) = 4 BA;1 B (5:6:23) 2n ; BA p 2n + (l )2n : By (5.6.20) and (5.6.23), we get the following relation for the gradient of the functional F grad F ( ) = 1 BA;1 B ; BA;1 p + l = ;Bq + l: (5:6:24) 2 Let PM be the orthogonal projection of R2n onto M , which is de ned by the formula =(
1 2

:::

2n 2n ) 2 R ( i +=

PM =

if i 0 if i < 0:

+=; + + 1 2

:::

+ 2n

To solve the problem (5.6.22), (5.6.23), we apply the gradient projection method. Taking an arbitrary initial approximation (0) 2 M , by using (5.6.24) and (5.6.20), we construct the sequence f (n) g M by the following formulas: 1 (n+1) = P q(n) = A;1 p ; 2 B (n) (5:6:25) M (n) ; n l ; Bq (n) n being a positive constant. The sequence fq(n) (n) g constructed by the algorithm (5.6.25) with a con1 1 stant n satisfying the condition 0 < n < d , where d = 2 kBA;1 B k, can be shown to converge to the solution (t ) of the problem (5.6.19).

Numerical solutions

The results below are obtained for the simply supported, rectangular plate = (x y) j ; a < x < a ;b < y < b of constant thickness h = 0:01 m, with a = b = 5h, made of an isotropic material for which the elasticity modulus E = 1:96 1011 Pa and Poisson's ratio = 0:3. There was taken f0 = const = 9:8 104 Pa and 16 control loads were chosen (n = 16)

5.6. Control by the load for plates and shells

335

which acted on the areas i (see (5.6.9), (5.6.10)). The location of the rst four areas i (see (5.6.8)) was de ned by the relations a1 = a2 = b1 = b3 = 10;2 m, a3 = a4 = b2 = b4 = 3 10;2 m, other areas are located symmetrically with respect to the x and y axes, the sizes of the areas being equal i = i = 10;3 m, i = 1 2 : : : 16. The optimization problem was solved by the algorithm (5.6.25). In the rst case, the goal functional was chosen in the form (M1 (wf0 +f ))2 + (M2 (wf0 +f ))2 + (M3 (wf0 +f ))2 dx dy ;a ;b (5:6:26) where M1 , M2 , and M3 are the bending moments and torque de ned by the formulas (4.1.3) and (4.1.4). M1 The results of calculation of the dimensionless moments M1 = 4f0 a2 and M3 M3 = 4f0 a2 are shown on Fig. 5.6.1. Here, the dashed lines correspond to the left ordinate scale, and they show the values of M1 along the line x = 0 (Curve 1), a and the values of M3 along the line x = ; 2 (Curve 2) provided the control loads are absent, i.e., when qi = 0.

J (f ) =

ZaZbh

* M1* , M 3

* M1* , M 3

0.04 0.03 0.02 0.01 0 - 0.01 6 5 2

1 4 3

0.004 0.003 0.002 0.001 0 - 0.001 - 0.002 0.6 y / b

- 0.02 - 1 - 0.6 - 0.2 0.2

Figure 5.6.1: Bending moments and torques under the optimal control load and in the absence of control Curves 3{6 correspond to the right ordinate scale, Curves 3 and 5 correspond to the solution of the optimization problem without restrictions, i.e., when qi; = ;1 and qi+ = 1 in (5.6.14), and they show the values of M1 and M3 along the lines x = 0 and x = ; a , respectively. Curves 4 and 6 correspond to the solution 2 of the optimization problem under the restrictions 0 qi qi+ , and they show the values of the moments M1 and M3 along the same lines x = 0 and x = ; a , 2

336

Chapter 5. Optimization of deformable solids

respectively. We note that Curves 1, 3, 4 are symmetric with respect to the point y = 0, while Curves 2, 5, 6 are antisymmetric. Fig. 5.6.1 shows that the values of M1 and M3 can be lowered by a factor 0.1 and less if the optimal values of the control loads are chosen. In the second case, the goal functional was chosen in the form

J (f ) =

ZaZb

;a ;b

2 wf0 +f dx dy

(5:6:27)

and the optimization problem was solved for the same input data and for the same location of the control loads. The results of calculation of the function of the dimensionless de ection

wf + 3 w = 192(10;f Ehf a4 2) 0
are shown in Fig. 5.6.2.
w* 105 w* 105

400 300 200 100 0

- 100

1 2 6 3 4 8 7 5

4 3 2 1 0 - 1 - 2 0.6 0.8 x , y a b

- 200 0 0.2 0.4

Figure 5.6.2: De ections under the optimal control load and in the absence of control Since the function wf0 +f is symmetric with respect to the x and y axes, only the values of w for 0 x a, 0 y b = a are presented. The dashed lines 1, 2 correspond to the left ordinate scale, and they show the values of w along the lines x = 0 and x = a , respectively, for the case when the control loads are absent, 2 i.e., when all qi are equal to zero. Curves 3{8 correspond to the right ordinate

5.7. Optimization of structures of composite materials

337

scale. Curves 3, 4 show the values of w along the lines x = 0 and x = a for the 2 solution of the optimization problem without restrictions, i.e., when qi; = ;1, qi+ = 1. Curves 5{8 picture the values of w along the lines x = 0, x = a , y = 0, 2 y = a , respectively, in the case of restrictions of the form 0 qi qi+ , where 2 qi+ are constants. In the latter case, since the restrictions are not symmetric with respect to the diagonal of the square, Curves 5, 7 and 6, 8 do not coincide. Fig. 5.6.2 shows that the maximal values of de ections of the plate can be reduced more than in hundred times by choosing the optimal values of control loads.

5.7 Optimization of structures of composite materials

\The music-master wrote a work in twenty- ve volumes, about the arti cial bird, which was very learned and very long, and full of the most di cult Chinese words yet all the people said they had read it, and understood it, for fear of being thought stupid and having their bodies trampled upon." | H. Ch. Andersen \The Nightingale"

5.7.1 Concept of a composite material

A material is called composite if it consists of two or more mutually insoluble components (phases). One of them is called a matrix, the other ones are called an inclusion, or a reinforcing material. Composites are classi ed according to the form of the inclusion. The most frequently used forms of an inclusion are granules (dispersible particles such that all dimensions of an arbitrary particle are of the same order), short (teared) bers, long bers ( laments), and laminates. By disposing long bers in parallel, one obtains a unidirectional ber reinforced composite laminate. It has a very high sti ness and strength in the ber direction, however in the transverse direction these characteristics are signi cantly worse. Collections of unidirectional ber reinforced composite laminates are widely used in practice. Glass, boron, and carbon bers are often applied, polymers and metals are used as a matrix. The dimensions of granules contained in a composite are negligible in comparison with the dimensions of the composite (an element of a structure made of the composite), and the composite contains a great number of granules.

338

Chapter 5. Optimization of deformable solids

In case of a ber reinforced composite, the diameters of bers are usually negligible as compared to the dimensions of the composite, and there are lots of bers in the composite. Therefore, homogenized (e ective) elastic or inelastic characteristics of a composite which are independent of points of the composite are applied. There are many approaches based on various physical or mathematical arguments which allow one to obtain homogenized characteristics of composites. We will dwell upon some of them.

5.7.2 Homogenization (averaging) of a periodical structure based on the G-convergence

For a periodical structure, e ective elasticity characteristics may be obtained by the construction of a G-limit operator. Below, we consider this problem, using the results of Duvaut (1976).

Problem of theory of elasticity for a periodical medium Suppose that an elastic medium occupies the whole space R3 . Denote by u = fuig the vector function of displacements of points of the medium, and by = f ij g,
i j = 1 2 3, the stress tensor function. Stresses and displacements in the medium
are connected by the Hooke law
ij =
3 X

k m=1

aijkm "km (u)

(5:7:1)

where "km (u) are the components of the strain tensor, 1 @u "km (u) = 2 @x k + @um m @xk

k m = 1 2 3:

(5:7:2)

We assume the coe cients aijkm to be Y -periodical functions on R3 , i.e., Y = 0 Y1 ] 0 Y2 ] 0 Y3 ] is the main period, and the functions aijkm de ned on Y are extended to R3 by periodicity. Suppose that

aijkm 2 L1(Y ) aijkm (x) = ajikm (x) = aijmk (x) = amkij (x) a.e. in Y
3 X

i j k m=1

aijkm (x) ij km c0
ij = ji 2 R

c0 = const > 0:

3 X 2 ij a.e. in Y i j =1

(5.7.3) (5.7.4) (5.7.5)

5.7. Optimization of structures of composite materials

339

a" (x) = aijkm x : (5:7:6) ijkm " Obviously, a" is a periodical function of x with the period "Y , satisfying the ijkm conditions (5.7.3){(5.7.5). Further, let be a bounded Lipshchitz domain in R3 , let S be the boundary of , S = S1 S2 , S1 \ S2 = ?, mes S1 > 0. De ne a set H by the formula ; H = fu j u = (u1 u2 u3 ) 2 W21 ( ) 3 ujS1 = 0g: (5:7:7)
The set H endowed with the scalar product and the norm of W21 ( ) 3 is a Hilbert space. De ne a bilinear form a" on H H by the relation
;

Let " be a positive number, which will be directed to zero, and let

a" (u v) =

aijkm x "km (u)"ij (v) dx: " i j k m=1

3 X

(5:7:8)

Suppose we are given vector functions f and F such that f = (f1 f2 f3 ) 2 (L2 ( ))3 F = (F1 F2 F3 ) 2 (L2 (S2 ))3 : (5:7:9) Denote by L the linear continuous functional in H determined by the formula (L v ) =
Z X 3

Now, consider the following problem: Find a function u" such that u" 2 H a" (u" v) = (L v) v 2 H: (5:7:11) Theorem 5.7.1 Let coe cients aijkm be Y -periodical functions on R3 , and let the conditions (5.7.3){(5.7.5) be satis ed. Assume that a bilinear form a" is determined by (5.7.8), and an element L 2 H by the formulas (5.7.9) and (5.7.10). Then, for all " > 0, the problem (5.7.11) has a unique solution, and the following estimate is valid ku" k(W21 ( ))3 const " > 0: (5:7:12) Proof. From (5.7.5) and (5.7.8), we conclude that a" (u u) c0 kuk2 u2H ">0 (5:7:13) H where Z X 3 2 = kukH ("ij (u))2 dx: (5:7:14)
i j =1

i=1

fi vi dx +

Z X 3

S2 i=1

Fi vi ds:

(5:7:10)

340

Chapter 5. Optimization of deformable solids

The Korn inequality implies that, in the space H , the formula (5.7.14) de nes ; a norm equivalent to the norm of W21 ( ) 3 . (Although in subsec. 1.7.2 we have proved the Korn inequality in the case R2 , the same technique will give us 3 .) the inequality in the case R Now, taking into account (5.7.3) and (5.7.13), from Theorem 1.5.2 we obtain that, for any " > 0, the problem (5.7.11) has a unique solution and the estimate (5.7.12) holds.

Homogenization of the structure

De ne a space W (Y ) as follows

W (Y ) = v j v 2 W21 (Y )

the traces of v are equal on the opposite sides of Y : (5.7.15)

Let us present another equivalent de nition of this space: W (Y ) is the restriction on Y of the vector functions that are de ned on R3 , Y -periodical, and ; the restrictions of which on any bounded open set Q R3 belong to W21 (Q) 3 .

Lemma 5.7.1 Let K0 be the set of constant three-dimensional vectors, and let W1 be the orthogonal complement of K0 to W (Y ) with respect to the scalar product of ; 1 W2 (Y ) 3 . Then, in the space W1 , the formula
kuk1 =
Z X 3

Y i j =1

("ij

(u))2

dx

1 2

(5:7:16)

; determines a norm equivalent to the norm of W21 (Y ) 3 :

Proof. Let u 2 ;W21 (Y ) 3 and let "ij (u) = 0, i j = 1 2 3. Then, the function
u = (u1 u2 u3 ) has the form (see, e.g., Necas and Hlavacek (1970)) u1 = a1 ; b3 x2 + b2 x3 u2 = a2 ; b1 x3 + b3 x1 u3 = a3 ; b2 x1 + b1 x2
(5.7.17)

where ai and bi are constants. Moreover, if u 2 W (Y ), then (5.7.15) and (5.7.17) yield that u = (a1 a2 a3 ) 2 K0 . Hence, if v 2 W1 and "ij (v) = 0, i j = 1 2 3, then v = 0. Now, Lemma 5.7.1 is a consequence of the Korn inequality and Theorem 1.7.3. Introduce the bilinear form

a(u v) =

3 X

Y i j k m=1

aijkm (x)"km (u)"ij (v) dx:

(5:7:18)

5.7. Optimization of structures of composite materials

341

Further, let P ij be the vector function de ned on Y and taking values in R3 , the i-th component of which is equal to xj and the other ones are equal to zero, i.e., P ij = fPkij g3=1 Pkij = xj ki i j k = 1 2 3: (5:7:19) k The relations (5.7.3){(5.7.5) together with Lemma 5.7.1 imply that there exists a unique function X ij such that ; ; X ij 2 W1 a X ij v = a P ij v v 2 W1 i j = 1 2 3: (5:7:20) Theorem 5.7.2 Let coe cients aijkm be Y -periodical functions on R3 satisfying the conditions (5.7.3){(5.7.5), and let functions f and F meet the conditions (5.7.9). Then, u" ! u weakly in H as " ! 0 (5:7:21) " is the solution of the problem (5.7.11) and u is the unique solution of the where u following problem u2H A(u v) = (L v) v 2 H: (5:7:22) The bilinear form A is de ned by

A(u v) =

and satis es the conditions A(u u) c0 kuk2 u2H H A(u v) = A(v u) u v2H (5.7.25) where c0 is the positive constant from (5.7.5). ^ Remark 5.7.1 Denote by A" and A the operators generated by the bilinear forms " and A, which are de ned through the relations a (A" v w) = a" (v w) v w2H (5.7.26) ^ (Av w) = A(v w) v w 2 H: (5.7.27)

gijkm = Y Y Y 1 2 3

gijkm "km (u)"ij (v) dx i j k m=1 1 a ;X ij ; P ij X km ; P km

3 X

(5.7.23) (5.7.24)

Theorem 5.7.2 states that the sequence of operators fA" g G-converges to the ^ operator A as " ! 0 (see Section 1.13). In the case of a Y -periodical structure ^ being examined, the G-limit operator A has constant coe cients, which are called e ective (or homogenized, or averaged) elasticity coe cients (constants). These coe cients do not depend on the boundary conditions, on the domain , and on the load L = ff F g. Indeed, (5.7.18), (5.7.20), and (5.7.24) show ^ that the coe cients gijkm of the operator A depend only on peculiarities of Y periodicity, i.e., only on the coe cients aijkm (x), x 2 Y , and the domain Y .

342

Chapter 5. Optimization of deformable solids

Remark 5.7.2 Obviously, if a function X ij is a solution of the problem (5.7.20),

then the function Y ij = X ij + C , C being an arbitrary constant vector from K0 , is a solution of the problem

Y ij 2 W (Y )

; ; a Y ij v = a P ij v

v 2 W (Y ):

(5:7:28)

The coe cients gijkm can be determined by the formula

; 1 gijkm = Y Y Y a Y ij ; P ij Y km ; P km : 1 2 3

From the point of view of computation, the problem (5.7.28) seems to be easier than the problem (5.7.20). Apropos of the solution of the problems (5.7.20) and (5.7.28), see subsec. 5.2.4.

Remark 5.7.3 Denote by E1 the subspace of rigid displacements determined by ; 3

(5.7.17). Let E2 be the orthogonal complement in W21 ( ) of the subspace E1 . ; Suppose L is an element of the dual of W21 ( ) 3 , and (L u) = 0 for all u 2 E1 . Then, for H = E2 , the problem (5.7.11) has a unique solution and the proof of Theorem 5.7.2 stated below remains valid. Hence, the estimate (5.7.25) holds true not only for u from H de ned by (5.7.7), but also for any function u 2 E2 . Take a function u = (u1 u2 u3 ) in the form

u1 = a11 x1
3 X

u2 = 2a21 x1 + a22 x2 gijkm akm aij c0

u3 = 2a31 x1 + 2a32 x2 + a33 x3 : aij = aji 2 R:

Then, by (5.7.25), taking to notice (5.7.2) and (5.7.23), we get

i j k m=1
3 X 2 aij i j =1

Using the equalities (5.7.4), (5.7.18){(5.7.20), and (5.7.24), one can verify that the elasticity coe cients gijkm meet the following symmetry conditions

gijkm = gjikm = gijmk = gmkij :

Proof of Theorem 5.7.2. 1. Introduce the notation

3 X " aijkm ij = k m=1

x " (u" ) : " km " > 0:

(5:7:29)

From (5.7.2), (5.7.3), and Theorem 5.7.1 (see (5.7.12)), we conclude that
" ij L2 ( )

const

(5:7:30)

5.7. Optimization of structures of composite materials

" f ij g such that

343

From here and (5.7.12), we derive the existence of subsequences fu" g and

"!0 u" ! u u" ! u

" ij

ij

; weakly in W21 ( ) 3 strongly in (L2( ))3 weakly in L2( ):

(5.7.31) (5.7.32) (5.7.33) (5.7.34)

Taking into account (5.7.8), (5.7.29), (5.7.34), and passing to the limit in (5.7.11), we get
Z X 3

2. Let P = fPi g3=1 and Pi = 3=1 cik xk , cik being constants. (5.7.3){(5.7.5) i k and Lemma 5.7.1 yield the existence of a unique function such that

i j =1

ij "ij (v ) dx = (L v )
P

v 2 H:

(5:7:35)

2 W1

a( v) = a(P v) v 2 W1 : w" (x) = P (x) ; " x "

(5:7:36) (5:7:37)

Introduce the following function on R3 :

in this case, the function (x) is Y -periodically extended to R3 . Due to the Y -periodicity, the set of functions f ( x )g is bounded in (L2 ( ))3 . " Hence, (5.7.31) and (5.7.37) imply that

w" ! P

strongly in (L2( ))3 :

(5:7:38)

From the relations (5.7.18) and (5.7.36), taking into consideration the Y periodicity of the functions aijkm and as well as the linearity of P , we deduce that
Z
3 X

R3 i j k m=1

aijkm (x)"km (P (x) ; (x))"ij (h) dx = 0

h 2 D(R3 ) 3 :

From here and (5.7.37), making the change of variable, we get

Z
R3

aijkm x "km (w" (x))"ij (h(x)) dx = 0 " i j k m=1

If ' 2 D( ) and v 2 H , then (5.7.39) yields

3 X

h 2 D(R3 ) 3 : (5:7:39)
(5:7:40)

a" (w" 'v) = 0:

344

Chapter 5. Optimization of deformable solids

Take v = 'w" in (5.7.11) and v = u" in (5.7.40). Subtracting the second relation from the rst one, we get a" (u" 'w" ) ; a" (w" 'u" ) = (L 'w" ): (5:7:41) Taking to notice the notation (5.7.29), we can transform the equality (5.7.41) as follows:
Z
3 X " " @' ij wi @xj i j =1

Z X 3 @' aijkm x "km (w" )u" @x dx = fi wi" ' dx: i j " i=1 i j k m=1
3 X 3 X

(5.7.42)

aijkm x "km (w" ) " k m=1 are "Y -periodical, so that they weakly converge in L2( ) to the average on Y of
the function

The functions

aijkm (x)"km (P (x) ; (x)) k m=1 which will be denoted by Mij (P ). Now, upon (5.7.32){(5.7.34) and (5.7.38), passing
to the limit in (5.7.42), we obtain
Z X 3

3 X

i j =1
Z X 3

Z @' @' Pi @x ; Mij (P )ui @x dx = ij j j

3 X

i=1

fi Pi ' dx: @

(5:7:43)

Eliminating fi from (5.7.43) by (5.7.35), we have

i j =1
Z

ij Pi @x
3 X

@' ; M (P )u @' dx = Z ij i @x
j j

('Pi ) dx: ij i j =1 @xj

3 X

Hence, the integration by parts gives

'

i j =1

(; ij "ij (P ) + Mij (P )"ij (u)) dx = 0

3 X

' 2 D ( ):
(5:7:44)

Thus,

Let P rs be a vector function such that P rs = fPkrsg3=1 , Pkrs = xs kr . The k relation (5.7.2) implies that (5:7:45) "ij (P rs ) = 1 ( ir js + is jr ): 2

i j =1

ij "ij (P ) =

3 X

i j =1

Mij (P )"ij (u):

5.7. Optimization of structures of composite materials

345

Setting in (5.7.44) P = P rs and using (5.7.45) together with the equality = sr , we conclude rs
rs =
3 X

Here,

i j =1

Mij (P rs ) "ij (u):

(5:7:46)

rs rs Y1 Y2 Y3 Y k m=1 aijkm (x)"km (P (x) ; X (x)) dx and the function X rs is the solution of the following problem X rs 2 W1 a (X rs v) = a (P rs v) v 2 W1 :

Mij (P rs ) =

3 X

(5:7:47) (5:7:48)

By (5.7.45) and (5.7.48), the formula (5.7.47) can be rewritten as

Mij (P rs ) =

Y1 Y2 Y3 Y k m p q=1

3 X

; apqkm (x)"km P rs (x) ; X rs (x)

"pq P ij (x) ; X ij (x) dx: (5.7.49)

(5:7:50) from (5.7.18) and (5.7.49), we get (5.7.24). Moreover, (5.7.35) and (5.7.46) yield (5.7.22). 3. We have above de ned the function u by (5.7.32) and showed that it satis es the equation (5.7.22). The function u obviously depends on an element L 2 H , so that it will be denoted as u(L). Since the space H is separable, by the diagonal process we can choose from the sequence of bilinear forms fa"g a subsequence, which still will be denoted by fa" g, such that u" (Q) ! u(Q) weakly in H for all Q 2 G (5.7.51) G H G is dense in H : Here, u" (Q) and u(Q) are solutions of the problems u" (Q) 2 H a" (u" (Q) v) = (Q v) v2H u(Q) 2 H A(u(Q) v) = (Q v) v 2 H: (5.7.52) Thus, on the set G, which is dense in H , we can de ne a linear operator B such that BQ = u(Q): (5:7:53) (5.7.5), (5.7.8), (5.7.14), and (5.7.26) imply (A" v v) c0 kvk2 v 2 H " > 0: (5:7:54) H

Now, writing

grsij = Mij (P rs )

346 Hence,
;

Chapter 5. Optimization of deformable solids

Q A;1 Q " ;1 2 L(H H ). where A"

c0 A;1 Q 2 " H

Q2H ">0

(5:7:55) (5:7:56)

(5.7.51) and (5.7.53) yield

A;1 Q ! u(Q) = BQ "

weakly in H for all Q 2 G:

Taking into account (5.7.51), (5.7.53), (5.7.56), we pass to the limit in (5.7.55), then we get (Q BQ) c0 lim inf A;1 Q 2 c0 kBQk2 Q 2 G: (5:7:57) " H H This relation implies that kB k c;1 and the operator B can be extended 0 by continuity to the whole space H . Thus, kB kL(H H ) c;1 . Using (5.7.51), 0 (5.7.53), and the inequality kA;1kL(H H ) c;1 for all " (see (5.7.54) and Theorem " 0 1.5.2), we deduce that

u" (Q) ! BQ

weakly in H for all Q 2 H :

(5:7:58)

By the continuity, we establish that B (Q) = u(Q) for all Q 2 H , where u(Q) is a solution of the problem (5.7.52). Now, the inequality (5.7.57) can be strengthened as follows

Q2H : By virtue of (5.7.3), there exists a positive number c1 such that kA" kL(H H ) c1 for all ":
Hence, so that (5.7.55) implies
;

(Q BQ) c0 kBQk2 H

(5:7:59)

kA;1 QkH c;1 kQkH " 1

c0 c;2 kQk2 H 1 Q 2 H " > 0: Q2H :

Q A; 1 Q "

This formula together with (5.7.58) yields (Q BQ) c0 c;2 kQk2 H 1 of (5.7.59), we have Thus, the operator B is coercive. Hence, there exists the inverse operator ^ B ;1 2 L(H H ), and by (5.7.27), (5.7.52), and (5.7.53) B ;1 = A. Now, by virtue

^ (Au u) c0 kuk2 u2H H which is equivalent to (5.7.25). Hence, the problem (5.7.22) has a unique solution and (5.7.58) holds true for any sequence of operators fA" g as " ! 0.

5.7. Optimization of structures of composite materials

347

^ Further, let u and v be arbitrary elements from H , and A" u" = Au, A" v" = ^ Av. Then, u" ! u and v" ! v weakly in H , and by the symmetry of the operators A" (see (5.7.4), (5.7.8), and (5.7.26)) we conclude that ^ ^ ^ ^ (Au v) = lim(Au v" ) = lim(A" u" v" ) = lim(A" v" u" ) = lim(Av u" ) = (Av u): ^ Thus, the operator A is symmetric, and the theorem is proved.

Remark 5.7.4 We have above considered the approach due to Duvaut (1976) to

the calculation of the G-limit operator for a periodical elastic structure. It is based on the extraction of converging subsequences and passage to the limit. Another approach to the calculation of the G-limit operator for a periodical structure is based on the two-scale expansion. The latter is widely used and has many applications, see Sanchez-Palencia (1980), Bakhvalov and Panasenko (1984), Oleinik et al. (1992). Many methods and problems of homogenization are examined by Bensoussan et al. (1978), Zhikov et al. (1993).

Remark 5.7.5 In the case of a periodical, perforated, elastic structure, the con-

dition (5.7.5) is not satis ed, because the functions aijkm vanish in holes. However, the homogenized elastic characteristics of this structure are de ned by the same formula (5.7.24) with X ij 's the solutions of the problem (5.7.20). However, from the conditions aijkm = 0 in the holes it follows now that zero surface forces are given on the boundary of the holes. For the proofs, see Oleinik et al. (1990).

With the results of Theorem 5.7.2, one can determine the e ective elasticity constants of a composite, i.e., the coe cients gijkm of the G-limit operator, see (5.7.24), assuming a corresponding structure to be Y -periodical. For example, in the case of a granule reinforced composite, one can choose, as a model, the structure shown in Fig. 5.7.1. The dashes show the space inside of Y that is occupied by a reinforcing element (granule). For a composite with oriented short bers, the structure shown in Fig. 5.7.2 can be taken as a model. The dashes show the space inside of Y that is occupied by a short ber. At last, for a unidirectional ber reinforced composite, the model of the structure can be taken as shown in Fig. 5.7.3. A model of unidirectional ber reinforced composite of a more complicated structure is shown in Fig. 5.7.4. There exist a number of distinct approaches giving the e ective elasticity constants of composites, which are based upon various physical reasonings (see, e.g., Van Pho Phy (1971b), Sendeckyj (1974), Pobedria (1984)). Below, given are some simple formulas used to determine the e ective elasticity constants of composite materials.

5.7.3 E ective elasticity characteristics of granule and ber reinforced composites

348

Chapter 5. Optimization of deformable solids

Y3

Y1

Y2

Figure 5.7.1: Structure of a granule reinforced composite

Y3

Y1

Y2

Figure 5.7.2: Structure of a composite with oriented short bers

5.7. Optimization of structures of composite materials

349

Y3

Y1

Y2

Figure 5.7.3: Structure of a unidirectional ber reinforced composite

Y3

Y1

Y2

Figure 5.7.4: Complicated structure of a unidirectional ber reinforced composite

350

Chapter 5. Optimization of deformable solids

In what follows, we will refer to the stresses de ned by the Hooke law (5.7.1) in which the functions aijkm are replaced by the e ective elasticity constants as macroscopic, or averaged, or homogenized stresses. In case of granule reinforced composites which contain spheroidal isotropic inclusions that are uniformly distributed in an isotropic matrix, the material remains macroscopically isotropic, i.e., the relations between the macroscopic stresses ij and the strains "ij (u) have the form (5:7:60) ij = e(u) ij + 2G"ij (u): Here,

e(u) =

3 X

which are determined by the relations (see Sendeckyj (1974)) = v1 1 + v2 2 G = v1 G1 + v2 G2 (5:7:62) where 1 , G1 and 2 , G2 are the elasticity constants of the inclusion and matrix, respectively, v1 and v2 are the volume fractions of the inclusion and matrix, v1 + v2 = 1: (5:7:63) The e ective elasticity constants and G can also be determined by the following formulas (Sendeckyj (1974)) 1 = v1 + v2 1 = v1 + v2 : (5:7:64) We point out that the relations (5.7.62) and (5.7.64) are consistent with the mixture rules. In case of a unidirectional ber reinforced composite (layer), the material is macroscopically orthotropic. If the x1 axis coincides with the direction of the bers (see Fig. 5.7.5), then the relations between the macroscopic stresses ij and the strains "ij (u) look as follows (see Sendeckyj (1974)) 11 = E11 "11 (u) + E12 "22 (u) (5.7.65) 22 = E21 "11 (u) + E22 "22 (u) 12 = 2G"12 (u): Here, The e ective elasticity constants E1 , E2 , 1 , 2 , and G are de ned by the formulas E1 = v1 Ef + v2 Em 1 = v1 f + v2 m
1 2

"ij (u) are de ned by the formulas (5.7.2), and G are the e ective Lame constants,

i=1

"ii (u)

(5:7:61)

G1

G2

Eii = 1 ;Ei

1 2

i=1 2

E12 = E21 = 2 E11 = 1 E22 :

(5:7:66)

5.7. Optimization of structures of composite materials

x2

351

x1

Figure 5.7.5: Unidirectional ber reinforced composite x1 axis coincides with the direction of the bers 1 v1 v2 E2 = Ef + Em 1 E2 = 2 E1 1 v1 v2 G = Gf + Gm v1 + v2 = 1: (5.7.67)

Here v1 and v2 are the volume fractions of the ber and matrix, Ef , Em , Gf , Gm , f , and m are the elasticity modules, shear modules, and the Poisson ratios of the ber and matrix. As noted by Sendeckyj (1974), the formulas (5.7.67) are in good agreement with experimental data. From (5.7.66) and (5.7.67) one sees that, for given materials, when Ef , Em , Gf , Gm , f , and m are xed, the e ective elasticity constants Eij , Ei , G, and i depend on the volume fraction v1 of the bers. Thus, further we will use the notations

G(v) = G Ei (v) = Ei i (v) = i : (5:7:68) Notice that, due to (5.7.66), E12 (v) = E21 (v), and (5.7.67) implies v2 = 1 ; v. Now, let the bers be directed at angle to the x1 axis (Fig. 5.7.6). Then, the relations between the macroscopic stresses and strains in the x1 and x2 coordinate
axes are of the form (see Malmeister et al. (1980))
11 = B11 ( v )"11 (u) + B12 ( v )"22 (u) + 2B16 ( v )"12 (u) (5.7.69) 22 = B21 ( v )"11 (u) + B22 ( v )"22 (u) + 2B26 ( v )"12 (u) = 21 = B61 ( v)"11 (u) + B62 ( v)"22 (u) + 2B66 ( v)"12 (u): 12

v = v1

Eij (v) = Eij

Here, the elasticity constants Bij ( v) are the components of a tensor of rank 4, which are de ned by the formulas

B11 ( v) = E11 (v) cos4 + E22 (v) sin4 + (4G(v) + 2E12 (v)) sin2 cos2 B12 ( v) = B21 ( v)

352

Chapter 5. Optimization of deformable solids

= (E11 (v) + E22 (v) ; 4G(v)) sin2 cos2 + E12 (v)(cos4 + sin4 ) B22 ( v) = E11 (v) sin4 + E22 (v) cos4 + (4G(v) + 2E12 (v)) sin2 cos2 B16 ( v) = B61 ( v) = sin cos (E11 (v) ; E12 (v)) cos2 ; (E22 (v) ; E21 (v)) sin2 ; 2G cos 2 ] B26 ( v) = B62 ( v) = sin cos (E11 (v) ; E12 (v)) sin2 ; (E22 (v) ; E21 (v)) cos2 + 2G cos 2 ] B66 ( v) = (E11 (v) + E22 (v) ; 2E12 (v)) sin2 cos2 + G(v) cos2 2 : (5.7.70)

x2

x1

Figure 5.7.6: Composite with bers directed at angle to the x1 axis Now, let us consider a composite formed by two identical unidirectional ber reinforced layers such that the directions of the bers of the layers form with the x1 axis angles and ; (Fig. 5.7.7). (5.7.69) and (5.7.70) yield this composite to be orthotropic, so that the relations between macroscopic stresses and strains in the x1 and x2 coordinate axes are of the following form
11 = B11 ( v )"11 (u) + B12 ( 22 = B21 ( v )"11 (u) + B22 ( 12 = 21 = 2B66 ( v )"12 (u)

v)"22 (u) v)"22 (u)

(5.7.71)

where the coe cients Bij ( v) are de ned by (5.7.70).

5.7. Optimization of structures of composite materials

x2

353

-q

x1

Figure 5.7.7: Composite formed by two identical ber reinforced layers the directions of the bers form with the x1 axis angles and ;

5.7.4 Optimization of the e ective elasticity constants of a composite

Setting of the optimization problem
Theorem 5.7.2 gives a method of calculation of the e ective elasticity constants for a Y -periodical structure. Now, we will consider the optimization problem when a Y -periodical structure is a control. Let the main period Y = 0 Y1 ] 0 Y2 ] 0 Y3 ] contain an inclusion. Denote by G the closed domain occupied by the inclusion, and by S its boundary. Assume that the elasticity properties of the inclusion and matrix are characterized by the elasticity constants ain and am (see (5.7.1)) which do not ijkm ijkm depend on the coordinates. Then, in the main period Y , the coe cients aijkm are de ned by the formula
in aijkm (x) = aijkm if x 2 G am if x 2 Y n G: ijkm

(5:7:72)

G is supposed to be a star-shaped domain with respect to the point x0 located at the center of Y , x0 = ( Y21 Y22 Y23 ). Denote by S0 the unit sphere in R3 centered at x0 . Then, the boundary S of the domain G is de ned by a continuous positive function f (s) of a point s 2 S0
(5:7:73) n(s) being the unit vector which is normal to the surface S0 at the point s. Let C (S0 ) be the space of real-valued continuous functions on S0 equipped with the norm kukC (S0) = max ju(s)j: (5:7:74) s2S
0

through the formula

x = x0 + f (s)n(s)

354

Chapter 5. Optimization of deformable solids

De ne a set U by the following relation

U = f j f 2 C (S0 ) sup 0

s s 2S0

hP

jf (s) ; f (s0 )j c0 i 3 (x (s) ; x (s0 ))2 1=2 i i=1 i

f1 (s) f (s) f2 (s) s 2 S0 : (5.7.75)

(5:7:76) (5:7:77)

Here, xi (s) = x0 + ni (s), ni (s) being the coordinate of the vector n(s), and i

f1 f2 2 C (S0 ), c1
positive constants.

f1(s)

f2 (s), s 2 S0 , c0 andc1 are s 2 S0 :

The function f2 is assumed to satisfy the following condition

x0 + f2 (s)n(s) 2 Y

By the formula (5.7.73), to each function f 2 U there corresponds the domain G(f ) occupied by the inclusion:

G(f ) = x j x = x0 + f (s)n(s) 0

< 1 s 2 S0 :

(5:7:78) (5:7:79)

The boundary S (f ) of the domain G(f ) is given by the formula

S (f ) = x j x = x0 + f (s)n(s) s 2 S0 :
(

Further, in accordance with (5.7.72), the coe cients af are de ned on the ijkm main period Y by the relation

af (x) = ijkm
Z

ain ijkm if x 2 G(f ) am ijkm if x 2 Y n G(f ) af (x)"km (u)"ij (v) dx: ijkm

(5:7:80)

and in the space W1 (see Lemma 5.7.1) the following bilinear form is well-de ned

af (u v ) =

3 X

Y i j k m=1

(5:7:81)

One can easily verify the following statement.

Lemma 5.7.2 Let ain and am be constants satisfying the conditions of symijkm ijkm
metry (5.7.4) and positive de niteness (5.7.5). Assume the coe cients af to ijkm be de ned by the relations (5.7.78) and (5.7.80). Then, for any f 2 U , the bilinear form af determined by the formula (5.7.81) is symmetric, continuous, and coercive in W1 .

5.7. Optimization of structures of composite materials

355

Lemma 5.7.2 implies that, for any f 2 U , there exists a unique function X ij (f ) such that ; ; X ij (f ) 2 W1 af X ij (f ) v = af P ij v v 2 W1 i j = 1 2 3 (5:7:82) P ij being de ned by (5.7.19). By virtue of Theorem 5.7.2, to every function f 2 U there correspond the e ective elasticity constants gijkm (f ) determined by the formula (5.7.83) Now, considering the function f as a control, we de ne the set of admissible controls by Uad = f j f 2 U k (f ) 0 k = 1 2 : : : l : (5:7:84) Here, 9 f ! k (f ) is a continuous mapping of U equipped with= (5:7:85) the topology generated by the one of C (S0 ) into R, k = 1 2 : : : l. If the functional 1 is chosen in the form
1 (f ) =

; 1 gijkm (f ) = Y Y Y af X ij (f ) ; P ij X km (f ) ; P km 1 2 3

i j k m = 1 2 3:

G(f )

dx ; c

(5:7:86)

c being a positive constant, then 1 de nes a restriction on the volume of inclusion, which can be dictated by the weight of the composite or its cost. Since ( Z Z dx = y(x) dx y(x) = 1 if x 2 G(f ) 0 if x 2 Y n G(f ) G(f ) Y
the Lebesgue theorem implies the functional (5.7.86) to meet the condition (5.7.85). Let a goal functional be of the form (f ) =
3 X

i j k m=1

(gijkm (f ) ; cijkm )2

(5:7:87)

cijkm being given constants.

The optimization problem consists in nding a function f0 such that f0 2 Uad (f0 ) = f 2U (f ): inf (5:7:88)
ad

356

Chapter 5. Optimization of deformable solids

(5.7.76), (5.7.84), (5.7.85), and let (5.7.77) hold. Assume that ain and am are ijkm ijkm constants satisfying the conditions of symmetry and positive de niteness (5.7.4), (5.7.5), and the coe cients af are determined by (5.7.78), (5.7.80). Further, ijkm let constants gijkm (f ) be de ned by the formulas (5.7.19), (5.7.81){(5.7.83) and let a goal functional be determined by (5.7.87). Then, the problem (5.7.88) has a solution. To prove Theorem 5.7.3, we need some auxiliary statements. Introduce spherical coordinates r, , ', which are connected with Cartesian coordinates by the formulas x1 = x0 + r sin cos ' 1 x2 = x0 + r sin sin ' (5.7.89) 2 0 + r cos x3 = x3 where r 0, 0 , and 0 ' < 2 . Then a point s of the unit sphere S0 is determined by the and ' coordinates (r = 1), i.e., s = ( '). Let Q = ( ') j 0 < < 0 < ' < 2 : According to (5.7.79), the function Pf : Q ! S (f ) de ned by the relations ' ! Pf ( ') = fPfi ( ')g3=1 i 1 ( ') = x0 + f ( ') sin cos ' Pf 1 2 ( ') = x0 + f ( ') sin sin ' Pf (5.7.90) 2 3 ( ') = x0 + f ( ') cos Pf 3 is a map of the surface Sf . First, assume that f 2 U and f is a continuously di erentiable function on S0 . (We remind that the notion of a di erentiability of a function de ned on a smooth manifold is introduced via local maps see, e.g., Lions and Magenes (1972).) Then, the surface measure ds is well-de ned on Sf . The surface measure of Sf n Pf (Q) is equal to zero, and so by the theorem on the area of a parametric manifold (see, e.g., (Schwartz (1967)), we obtain
Z

The existence theorem Theorem 5.7.3 Let a nonempty set Uad be de ned by the relations (5.7.75),

Sf

ds =

ZZ

Df ( ') d d'
!2
3 i X @Pf

(5:7:91)
2 1 2

where

Df ( ') =

3 X

i=1

@Pfi @

!2 3 X

i=1

@Pfi @'

i=1

@Pfi @ @'

(5:7:92)

5.7. Optimization of structures of composite materials

357

If f is an arbitrary function from U , then, due to (5.7.75), f satis es the Lipschitz condition with the constant c0 . Thus, the partial derivatives of the function f exist a.e. in Q, and the following estimates are valid

@f @

c0 c2 @Pf @'

@f @'

c0 a.e. in Q: c2 a.e. in Q f 2 U:
(5:7:93)

This relation together with (5.7.90) implies For any u 2 D(Q), by applying the integration by parts formula (e.g., Schwartz (1967)), we obtain

@Pf @

@u d d' = ; ZZ Pf @ Q Q ZZ ZZ @u Pf @' d d' = ;

ZZ

1 Since Pf 2 W1 (Q) 3 , there exists a sequence of smooth functions fPn g meeting the conditions

@Pf u d d' @ @Pf u d d': @'

@Pn (' ) @

@Pn ! @Pf a.e. in Q @ @ @Pn ! @Pf a.e. in Q (5.7.94) @' @' @Pn (' ) c c2 (' ) 2 Q 8n: 2 @'

The relations (5.7.92) and (5.7.94) together with the Lebesgue theorem justify the passage to the limit as n ! 1 in (5.7.91) provided Pn is substituted for Pf . Finally, the formulas (5.7.91) and (5.7.92) de ne the measure of the surface Sf for any f 2 U , which does not depend on a choice of local maps, just as in the case of a surface of the C 1 class. Moreover, by virtue of (5.7.93), the measures of the surfaces Sf are bounded for all f 2 U . Thus, we have proved the following Lemma 5.7.3 Let a set U and a surface Sf be de ned by the relations (5.7.75), (5.7.76), and (5.7.79). Then, the relations (5.7.90){(5.7.92) determine the measure of the surface Sf , and there exists a number c3 such that
Z

Sf

ds c3

f 2 U:

(5:7:95)

358

Chapter 5. Optimization of deformable solids

In what follows, we will need also the following lemma.

Lemma 5.7.4 Let a set U be de ned by the relations (5.7.75) and (5.7.76), and

let ain and am be constants satisfying the conditions of symmetry and positive ijkm ijkm de niteness (5.7.4) and (5.7.5). Assume that a bilinear form af is de ned by the relations (5.7.78), (5.7.80), and (5.7.81), and a function u(f g) is a solution of the following problem

u(f g) 2 W1

af (u(f g) v) = (g v)

v 2 W1

(5:7:96)

where g 2 W1 . Then, f ! af is a continuous mapping of U , equipped with the topology induced by the C (S0 )-one, into L2 (W1 W1 R), and the function f g ! u(f g) is a continuous mapping of U W1 , endowed with the topology induced by the product of the topologies of C (S0 ) and W1 , into the space W1 :

Proof. Suppose that ffng is a sequence such that fn 2 U 8n fn ! f0 in C (S0 ): (5:7:97) Then, f0 2 U and taking into account (5.7.80), (5.7.81), and (5.7.95), for any u v 2 W1 , we have afn (u v) ; af0 (u v) c4 kfn ; f0 kC (S0) kukW1 kvkW1 :
This and (5.7.97) yield

afn ! af0

in L2 (W1 W1 R):

(5:7:98)

Denote by A(f ) the operator generated by the bilinear form af : (A(f )u v) = af (u v)

u v 2 W1 :

(5.7.97) and (5.7.98) imply that f ! A(f ) is a continuous mapping of U into L(W1 W1 ). Now, Theorem 1.8.1 implies that the function f g ! u(f g) is a continuous mapping of U W1 into W1 . Proof of Theorem 5.7.3. Let ffng be a minimizing sequence:

fn 2 Uad 8n

n!1

lim (fn ) = f 2U inf

ad

(f ) :

(5:7:99)

By virtue of (5.7.75), the functions fn satisfy the Lipschitz condition with the constant c0 . Thus, the generalized Arzela theorem (see, e.g., Kolmogorov and Fomin (1975)) implies that from the sequence ffng one can extract a subsequence ff g such that f ! z in C (S0 ): (5:7:100)

5.7. Optimization of structures of composite materials

359

(5.7.75), (5.7.84), (5.7.85), and (5.7.100) yield that z 2 Uad . Further, Lemma 5.7.4 and (5.7.100) imply

X ij (f ) ! X ij (z )

in W1 as ! 1 i j = 1 2 3

(5:7:101)

where X ij (f ) and X ij (z ) are the solutions of the problem (5.7.82) for f equal to f and z . From Lemma 5.7.4 and (5.7.100), we deduce that

af ! az

in L2 (W1 W1 R): (5:7:102)

Hence, (5.7.83), and (5.7.101) imply lim gijkm (f ) = gijkm (z ): !1 Now, (5.7.87), (5.7.99), and (5.7.102) yield (z ) = f 2U inf
ad

(f ):

Thus, the function f0 = z is a solution of the problem (5.7.88) and the theorem is proved.

Then, Theorem 5.7.3 remains valid provided the functional the functional from (5.7.87).

Remark 5.7.6 Denote by n the number of the e ective elasticity constants gijkm , and let g ! (g) is a continuous mapping of Rn into R. On the set U de ned by (5.7.75) and (5.7.76), introduce a function 0 by 0 (f ) = (g(f )), where g(f ) = fgijkm (f )g, and the functions f ! gijkm (f ) are determined by (5.7.83).
0

is substituted for

Let be a bounded domain in R3 occupied by a granule reinforced composite, and let S be the boundary of , S = S1 S2 , S1 \ S2 = ?, mes S1 > 0. The composite is fastened on S1 , while the part S2 is subject to a load. Making the assumption that the composite contains spherical inclusions distributed in an isotropic matrix, we will use the formulas (5.7.60){(5.7.63) to determine the macroscopic stresses in the material. According to (5.7.60){(5.7.63), the strain energy generates the following bilinear form

5.7.5 Optimization of a granule reinforced composite

A(u v) =
=

Z X 3 Z

i j =1

ij (u)"ij (v ) dx
3 X

e(u)e(v) + 2G

i j =1

"ij (u)"ij (v) dx:

(5.7.103)

360

Chapter 5. Optimization of deformable solids

Here, and G are the e ective elasticity constants. Introducing the notation g = v1 and taking notice of (5.7.62) and (5.7.63), we get = g 1 + (1 ; g) 2 G = gG1 + (1 ; g)G2 : (5:7:104) Suppose that the volume fraction of the inclusion, g, is a function of x. This assumption requires indicating the way of calculation of the function g(x) provided the structure of the composite (the distribution of the inclusion and matrix over the domain ) is known. Given an averaging radius that is su ciently large as compared to the dimensions of the granules, the function g(x) can be determined by

g(x) =

where ! is the averaging kernel (see subsec. 1.6.4), and the function (y) equals one if there is an inclusion at the point y, and vanishes otherwise. The value is considered to be su ciently small as compared to the dimensions of the composite. To calculate g(x) at any point x 2 by the formula (5.7.105), the function (y) should be extended to a domain larger than . Suppose we have made such an extension. Then, the function g(x) is smooth and even in nitely di erentiable in if so is the function ! . The formulas (5.7.104) determining the e ective elasticity constants are obtained by averaging over a su ciently large volume under the supposition that the inclusions are uniformly distributed over the volume of the matrix. (We stress that, in Theorem 5.7.2, the e ective elasticity constants were obtained under the condition of the Y -periodicity, which in case of a granule reinforced composite corresponds to a uniform distribution of inclusions over the volume of the matrix). Hence, in order to apply the formulas (5.7.104) in the case when g is a function, the latter must be smooth and its partial derivatives at every point must be small in absolute value, since increase of the absolute values of the derivatives leads to decrease of accuracy of the formulas (5.7.104). Thus, in what follows, in examining the optimization problem for the granule reinforced composite, the function g is supposed to belong to the set Ylc of the form Ylc = g j g 2 W2l ( ) kgkW2l( ) c1 c2 g(x) c3 (5:7:106) where c = (c1 c2 c3 ) c1 > 0 0 < c2 < c3 < 1 l 2: (5:7:107) Introducing the notation Ag = A and taking to notice (5.7.103) and (5.7.104), we get

R3

! (x ; y) (y) dy

(5:7:105)

Ag (u v) =

g 1 + (1 ; g) 2 ]e(u)e(v)

5.7. Optimization of structures of composite materials

361
o

+ 2 gG1 + (1 ; g)G2 ]

3 X

i j =1

"ij (u)"ij (v) dx: 5.7.108) (

The bilinear form Ag is de ned on the space H determined by (5.7.7).

Theorem 5.7.4 Let 1, 2 , G1, and G2 be positive numbers, g 2 L1( ), 0 g 1 a.e. in , let a bilinear form Ag be de ned by (5.7.108), and let a space H be determined by (5.7.7). Then, for any L 2 H , there exists a unique function u(g L) such that
u(g L) 2 H Ag (u(g L) v) = (L v) v 2 H:
(5:7:109)

Proof. The bilinear form Ag is symmetric and continuous, and by the Korn in-

equality it is coercive in H . Hence, the existence of a unique function u(g L) satisfying the conditions (5.7.109) is implied by the Riesz theorem. Further, assume that the composite is subject to the load L 2 H de ned by the relations (5.7.9) and (5.7.10), and the function g is a control. So, the solution of the problem (5.7.109) will be denoted by ug , i.e., ug = u(g L). Let there be given functionals k : Ylc H ! R such that

g u ! k (g u) is a continuous mapping of Ylc H equipped9 > with the topology generated by the product of the W2l ( )-= weak topology and the H -strong topology into R, k => 0 1 2 : : : p. Uad = g j g 2 Ylc
k (g ug )

(5:7:110)

Now, de ne a set of admissible controls as 0 k = 1 2 ::: p

0 ( g u g ):

(5:7:111) (5:7:112)

and let a goal functional be of the form (g ) = The optimization problem is to nd a function g0 such that

g0 2 Uad

(g0 ) = g2U inf

ad

(g):

(5:7:113)

The functionals k , k = 1 2 : : : p, can determine restrictions on the strength, sti ness, volume fraction of inclusion in the composite, and so forth. If the goal functional de nes the mass of the composite, then it is of the form (g) = ( Here,
in in ; m ) g dx +

dx:

(5:7:114)

and

are the densities of the inclusion and matrix.

362

Chapter 5. Optimization of deformable solids

If the functional

de nes a restriction on the sti ness, it can be chosen as

1 (g ug ) = Ag (ug

ug ) ; c1

(5:7:115)

Since the imbedding of Ylc into C ( ) is compact, the functional 1 from (5.7.115) can be shown to satisfy the condition (5.7.110). If the functional 2 determines a restriction on the volume fraction of the inclusion in the composite, then it is of the form
2 (g ug ) =

c1 being a positive number.

g dx ; c1

c1 = const > 0:

Theorem 5.7.5 Let 1 , 2, G1, and G2 be positive numbers, let a bilinear form Ag be de ned by (5.7.108), and let ug = u(g L) be the solution of the problem (5.7.109). Assume a nonempty set Uad is de ned by the relations (5.7.106),
(5.7.107), (5.7.110), and (5.7.111), and the goal functional by (5.7.112). Then, the problem (5.7.113) has a solution. Proof of Theorem 5.7.5 is based upon the fact that, for a three-dimensional bounded domain , the imbedding of W2l ( ) into C ( ) is compact for l 2. This proof is analogous to the one of Theorem 2.2.2, and we therefore omit it.

E ective relations of elasticity and the strain energy of a laminate shell

Consider a shell consisting of unidirectional ber laminates provided the distribution of the laminates is symmetric with respect to the midsurface. Moreover, next to every laminate the reinforcement direction of which forms an angle with the coordinate line, there is a layer with the reinforcement direction ; (see Fig. 5.7.7). Here, we assume that the coordinate lines on the midsurface coincide with the principal curvature lines of this surface. In accordance with results of subsec. 5.7.3, two identical layers which are located at angles and ; are treated as one layer in which the relations between the macroscopic stresses and strains are determined by the formulas (5.7.70) and (5.7.71). We will use the classical shell theory, see Section 4.5. Then, (5.7.71) and (4.5.4) imply that the stresses 11 , 22 , 12 at a point located at distance from the midsurface of the shell, which are denoted by 11 , 22 , 12 , are de ned by the following relations
11 = B11 ( ( ) v ( ))("1 + 1 ) + B12 ( ( 22 = B21 ( ( ) v ( ))("1 + 1 ) + B22 ( ( 12 = 21 = B66 ( ( ) v ( ))("12 + 2 12 ):

5.7.6 Optimization of composite laminate shells

) v( ))("2 + ) v( ))("2 +

2) 2 ) (5.7.116)

5.7. Optimization of structures of composite materials

363

Here, "1 , "2 , "12 , 1 , 2 , 12 are the components of the strains that are determined by the components of displacements of the midsurface of the shell through the formulas (4.5.5), and Bij ( ( ) v( )) are de ned by the relations (5.7.66){ (5.7.68), (5.7.70). (Notice that in (4.5.5) v is the component of the vector function of displacements of points ( ) of the midsurface, while in the formulas (5.7.69), (5.7.70) v is the function of volume fraction of the bers depending on .) To every 2 0 h n f i g, where h is the thickness of the shell and f i g is the set of points 2 belonging to the boundaries of the laminates on the intervals 0 h , we place in 2 correspondence ( )|the angle of orientation of the bers, i.e., the angle between the direction of the bers and the coordinate line, and the volume fraction of the bers v( ). By virtue of the above assumption on the symmetricity of the distribution of the laminates with respect to the midsurface, we have ( ) = (; )

v( ) = v(; ):

(5:5:117)

The strains "1 , "2 , and "12 at a point located at a distance from the midsurface are determined by the formulas (4.5.4), and in accordance with (4.5.7) and (4.5.8) the strain energy of the shell is given by
h 1 ZZ Z 2 ( =2 ;h 2

11 "1

22 "2

12 "12 ) d

AB d d :

(5:7:118)

Here, and are the curvilinear coordinates on the midsurface of the shell, A and B are the coe cients of the rst quadratic form. In (5.7.118), we substituted A and B for H1 and H2 , neglecting the values R1 and R2 as they are small as compared to unit, R1 and R2 being the principal curvature radii of the midsurface of the shell. Substituting in (5.7.118) the expressions of ij from (5.7.116) and the expressions of "1 , "2 , "12 from (4.5.4), taking to notice (5.7.117), and integrating in , we obtain ZZ =1 c11 (h v)"2 + 2c12 (h v)"1 "2 1 2 +1 2 where
ZZ

+ c22 (h

v)"2 + c66 (h v)"2 AB d d 2 12 D11 (h v) 2 + 2D12 (h v) 1 2 1 v) 2 + 4D66 (h v) 2

Z h 2
0 2 12

+ D22 (h

AB d d

(5.7.119)

cij (h v) = 2

Bij ( ( ) v( )) d

364

Chapter 5. Optimization of deformable solids

Dij (h v) = 2

Z h 2
0

Bij ( ( ) v( )) 2 d :

(5.7.120)

Thus, the strain energy of a laminate shell is de ned by the formula (5.7.119) in which the elasticity constants cij and Dij depend on the thickness of the shell as well as on the distributions of the angles of the orientation of the bers and their volume fraction.

Step and continuous controls

Let the midsurface of the shell go through the boundary of a laminate, and let n be the number of the laminates which are above the midsurface. Denote by bi the angle of the orientation of bers, and by gi the volume fraction of bers in the i-th laminate, i = 1 2 : : : n. Further, suppose that the laminates are of the same thickness l = 2h . Then, n ( ) and v( ) are the step functions of the following form ( ) = bi

v( ) = gi

for 2 ((i ; 1)l il) i = 1 2 : : : n:

(5.7.121)

On the boundaries of the laminates, i.e., at the points k = kl, k = 1 2 : : : , n ; 1, the functions and v are not well-de ned. Suppose the conditions (5.7.117) are satis ed. De ne sets G1 and G2 by the formulas

G1 = b j b = fbi g 2 Rn bi 2 0 2 ] i = 1 2 : : : n G2 = g j g = fgi g 2 Rn gi 2 e1 e2 ] i = 1 2 : : : n e1 and e2 being constants, 0 < e1 < e2 < 1. By (5.7.121), to every vector b 2 G1 there corresponds the function , and to every g 2 G2 the function v. We will denote these functions as b and vg . Introduce mappings Pij : G1 G2 ! R and Gij : G1 G2 ! R by the following
relations

By using (5.7.66){(5.7.68) and (5.7.70), one can easily verify that Pij and Gij are continuous functions on G1 G2 . Now, we can examine various problems of optimization of laminate shells in which optimization (control) parameters are the angles of the orientation of the bers, the volume fractions of bers in laminates, and the number of laminates. Since the thickness of a laminate is neglectibly small as compared to the thickness of the shell, the laminates can be considered to have zero thickness, and and v to be continues functions. The thickness of the laminate shell, h, which is independent of the coordinates of points of the midsurface of the shell, can be treated as a control.

Pij (b g) = cij (h b vg ) Qij (b g) = Dij (h b vg ) where cij and Dij are determined by (5.7.120).

(5:7:122)

5.7. Optimization of structures of composite materials

365

Now, de ne a set of controls by the formula

U = t j t = (q v) q 2 e3 e4 ]
where

v 2 C ( 0 e4 ]) 0

2 e1

v e2

(5.7.123)

e1 , : : : , e4 are positive numbers, e1 < e2 < 1: (5:7:124) For the triple (q v) 2 U , denote by q and vq the restrictions of the functions and v on the interval 0 q]. On the set U , let us de ne the functionals Rij (q v) = 2 Gij (q v) = 2
Zq Z0 q
0

Bij ( q ( ) vq ( )) d Bij ( q ( ) vq ( )) 2 d
(5.7.125)

Bij being de ned by the formulas (5.7.66){(5.7.68) and (5.7.70).

(5.7.120) and (5.7.125) imply that

cij (2q q vq ) = Rij (q v) Dij (2q q vq ) = Gij (q v): (5:7:126) Lemma 5.7.5 Let a set U be de ned by the relations (5.7.123) and (5.7.124), and functionals Rij and Gij by the formulas (5.7.125), (5.7.66){(5.7.68), and (5.7.70). Then, Rij and Gij are continuous mappings of U , equipped with the topology generated by the product of the topologies of R and (C ( 0 e4 ]))2 , into R.
One can easily prove Lemma 5.7.5 by using the Lebesgue theorem and the fact that the functions v ! Bij ( v) are continuous and bounded for all 2 0 2 ] and v 2 e1 e2 ]. Let us sum up the above arguments in the following remark.

Remark 5.7.7 To every index n|the number of the laminates above the mid-

surface of the shell|there correspond two sets in the n-dimensional space of step functions which determine virtual angles of the orientation of bers and the volume fractions of bers in the laminates. The thickness of the shell, being considered as a function of n, takes discreet values 2ln, where l is the thickness of the laminate. For small l and great n, one can consider the functions of orientation and of volume fraction of bers v as continuous functions, then the thickness of the shell also changes continuously. Such a passage from a discreet problem in n, with step functions of orientation and volume fraction, to a continuous problem is natural because any continuous function on an interval can be uniformly approximated by step functions, and visa versa, and because the functionals Rij and Gij are continuous with respect to the topology of uniform convergence (see Lemma 5.7.5).

366

Chapter 5. Optimization of deformable solids

However, for a given n (which can be rather great), in order to approximate a continuous function by a step function with a good accuracy, the continuous function should satisfy the Lipschitz condition. So, below, considering the optimization problem for shells, we suppose that the functions from the set of admissible controls satisfy the Lipschitz condition with some constant common for all functions. Below, to be more speci c, we will examine a cylindrical shell.

The strain energy of a cylindrical shell is determined by the relation (5.7.119). According to results of Section 4.6, setting = z , = ', we get A = 1, B = r = const and = (z ') j 0 < z < L 0 < ' < 2 where r is the radius of the midsurface of the shell and L is its length. The strain components are de ned by the formulas (4.6.5). Let t = (q v) 2 U . The strain energy, de ned by (5.7.119), generates a bilinear form at , which, in view of (5.7.126) and (4.6.10), has the form at (!0 !00 ) = =
Z LZ 2 n
0
0 00 0 00 0 00

Bilinear form and the theorem on the continuity of solutions for a laminate cylindrical shell

R11 (t)(P1 ! )(P1 ! ) + R12 (t) (P1 ! )(P2 ! ) + (P2 ! )(P1 ! ) 0 + R22 (t)(P2 !0 )(P2 !00 ) + R66 (t)(P3 !0 )(P3 !00 ) + G11 (t)(P4 !0 )(P4 !00 ) + G12 (t) (P4 !0 )(P5 !00 ) + (P5 !0 )(P4 !00 ) + G22 (t)(P5 !0 )(P5 !00 ) o + 4G66 (t)(P6 !0 )(P6 !00 ) r dz d': (5.7.127)

Let V be the Hilbert space de ned by the relations (4.6.8) and (4.6.11), and let f 2 V . Given t 2 U , consider the problem: Find a function !(t f ) such that !(t f ) 2 V at (!(t f ) !0) = (f !0 ) !0 2 V: (5:7:128) Theorem 5.7.6 Let a set U be determined by the formulas (5.7.123) and (5.7.124), and let a bilinear form at , t 2 U , be de ned by the relations (5.7.127), (5.7.66){(5.7.68), (5.7.70), and (5.7.125), provided the operators Pi in (5.7.127) are determined by the formulas (4.6.5) and (4.6.10) with A = 1 and B = r = const. Further, assume that the elasticity constants of the ber and matrix meet the following conditions : Ef Em Gf Gm are positive constants, (5.7.129) (5.7.130) f and m are constants, 0 f < 1, 0 m < 1. Then, for any (t f ) 2 U V , there exists a unique solution of the problem (5.7.128), and the function t f ! !(t f ) determined by this solution is a continuous mapping of U V , endowed with the topology generated by the product of the topologies of R, (C ( 0 e4 ]))2 and V , into the space V .

5.7. Optimization of structures of composite materials

367

Proof. (5.7.70) implies that, for any

1 2 6

2 R, 2 0 2 ], and v 2 e1 e2 ],
(5.7.131)

2 2 2 B11 ( v) 1 + 2B12 ( v) 1 2 + B22 ( v) 2 + 4B66 ( v) 6 = A1 ( v ) + A2 ( v )

where

1 2 2 A1 ( v ) = 2 B11 ( v) 1 + 2B12 ( v) 1 2 + B22 ( v) 2 2 + 4B66( v) 6 + 4B16 ( v) 1 6 + 4B26( v) 2 6 2 2 A2 ( v ) = 1 B11 (; v) 1 + 2B12(; v) 1 2 + B22 (; v) 2 2 2 + 4B66(; v) 6 + 4B16 (; v) 1 6 + 4B26 (; v) 2 6 : (5.7.132) By substituting Bij ( v) from (5.7.70) into (5.7.132), we get 1 A1 ( v ) = 2 E11 (v) 2 + 2E12 (v) 1 2 + E22 (v) 2 + 4G(v) 2 1 2 6 1 2 2 2 A2 ( v ) = 2 E11 (v) 1 + 2E12 (v) 1 2 + E22 (v) 2 + 4G(v) 6
2 + 2 + 2 2 = 2 + 2 + 2 2 = 2 + 2 + 2 2: 1 2 6 1 2 6 1 2 6

(5.7.133)

and

(5:7:134) From the physical point of view, the relations (5.7.133) mean that the strain energy does not depend on a speci c coordinate system. This is implied by the fact that the elasticity constants Bij ( v), de ned by (5.7.70), are tensors of rank 4, while the constants i , i , i in the formulas (5.7.132), (5.7.133) take the part of the components of the strain tensor respectively in the original coordinate system and in the coordinate systems rotated at angles and ; ), the rank of the strain tensor is equal to 2. The equality (5.7.134) corresponds to the fact that the sum of the squares of the strain components is invariant (see also Malmeister et al. (1980)). (5.7.67) and (5.7.68) yield the following formulas:

E1 (v) = vEf + (1 ; v)Em EE E2 (v) = vE + f(1 m v)E ; f m (v) = v f + (1 ; v) m 1 1 (v )E2 (v ) 2 (v ) = E (v ) :

1

(5.7.135) (5.7.136) (5.7.137) (5.7.138)

From (5.7.135) and (5.7.136) we obtain

E2 (v) = Ef Em 2 E1 (v) Ef Em (1 + 2v2 ; 2v) + (Ef2 + Em )(v ; v2 ) :

368

Chapter 5. Optimization of deformable solids

2 This, together with the inequality Em + Ef2 2Em Ef , implies that E2 (v) 1 v 2 e1 e2 ] E1 (v) e1 and e2 being constants satisfying the condition (5.7.124). (5.7.130) and (5.7.137) imply that max (v) < 1: v2 e e ] 1
1 2

(5:7:139)

(5:7:140) (5:7:141)

Further, (5.7.138){(5.7.140) yield max (v) 2 (v) < 1: v2 e e ] 1

1 2

From here and (5.7.135), (5.7.141), we obtain (5.7.143). Further, (5.7.66), (5.7.141), and (5.7.143) imply E11 (v)E22 (v) ; (E12 (v))2 c2 > 0 v 2 e1 e2 ]: (5:7:144) (5.7.142), (5.7.144), and the Sylvester criterion yield that the matrix Eij ], i j = 1 2, is positive de nite for any v 2 e1 e2], and, since v ! Eij (v) are continuous functions on e1 e2 ], there exists a constant c3 > 0 such that ; E11 (v) 2 + 2E12 (v) 1 2 + E22 (v) 2 c3 2 + 2 1 2 1 2 v 2 e1 e2] 1 2 2 R: Hence, (5.7.131){(5.7.134), (5.7.142), (5.7.125), and (5.7.127) imply that

Now, from (5.7.66){(5.7.68), (5.7.129), (5.7.135), (5.7.136), and (5.7.141), we deduce that 9 the functions v ! E11 (v), v ! E22 (v), and v ! G(v), being= (5:7:142) considered as mappings of e1 e2] into R, are continuous and bounded below and above by positive constants. Let us show that E1 (v) ; ( 1 (v))2 E2 (v) c0 > 0 v 2 e1 e2]: (5:7:143) If f and m vanish, then (5.7.143) is a consequence of (5.7.129), (5.7.135), and (5.7.137). Therefore, we assume that at least one of the constants f and m is not equal to zero. Then, from (5.7.130) and (5.7.137), we get 1 (v) c1 > 0 for any v 2 e1 e2 ]. Hence, (5.7.138) yields E E2 (v) = 2 (v)(v1)(v) :
1

at (! !) c4

Z LZ 2 X 6
0 0

i=1

(Pi !)2 dz d'

!2V t2U

(5:7:145)

5.7. Optimization of structures of composite materials

369

c4 being a positive constant.

The system of operators fPi g6=1 is coercive in V (see the proof of Theoi rem 4.6.1). Hence, (5.7.145) implies the existence of a number c5 > 0 such that

at (! !) c5 k!k2 V

! 2 V t 2 U:

(5:7:146)

The bilinear form at is continuous and symmetric in V . Lemma 5.7.5 yields that t ! at is a continuous mapping of U into L2 (V V R). Now, Theorem 1.8.1 yields that the function t f ! !(t f ), where !(t f ) is the solution of the problem (5.7.128), is a continuous mapping of U V into V . The theorem is proved.

Problem of optimization of a composite laminate cylindrical shell

Let a cylindrical shell be fastened in such a way that it does not have rigid displacements, i.e., displacements without deformations. From the mathematical point of view, this corresponds to the case when the space V de ned by the way of the fastening of the shell satis es the condition (4.6.11). In particular, if the edges are clamped, the space V is of the form

V = ! j ! = (u v w) 2 W !(0 ') = !(L ') = 0 @w (0 ') = @w (L ') = 0 (5.7.147) @z @z L being the length of the shell. (Notice that, in this relation, v is a component
of the vector function of displacements of the midsurface of the shell, but not the function of the volume fraction of the bers.) For the fastened edges, we have

V = ! j ! = (u v w) 2 W !(0 ') = !(L ') = 0 :

Z LZ 2
0 0

(5:7:148)

The shell is subject to a xed load, which is identi ed with an element f 2 V . In particular, f can be de ned by the following expression (f !) = (f1 u + f2 v + f3 w) r dz d'

fi 2 L2 ((0 L) (0 2 )): (5:7:149)

We suppose the assumptions of Theorem 5.7.6 to be ful lled. Then, for t 2 U , there exists a unique function !(t) such that

!(t) 2 V

at (!(t) !0 ) = (f !0 )

!0 2 V:

(5:7:150)

Here, we write !(t) instead of !(t f ), because the element f is xed, while t takes values in U .

370

Chapter 5. Optimization of deformable solids

Further, let there be de ned functionals k such that 9 t ! ! k (t !) is a continuous mapping of U V , equipped= with the topology generated by the product of the topologies of R, (C ( 0 e4 ]))2 , and V , into R, k = 0 1 2 : : : l. De ne a set of the admissible controls by

(5:7:151)

Uad = t j t = (q v) 2 U
0

sup jv(y) ; v(y )j c2 k (t !(t)) 0 k = 1 2 : : : l (5.7.152) 0 2 0 e4 ] jy ; y0 j yy and let a goal functional be of the form (t) = 0 (t !(t)): (5:7:153) The functionals k can de ne restrictions on the strength and sti ness of the shell, the volume fraction of bers in the composite, the cost of the composite, and so forth. As a condition on the fracture of a laminate composite, one frequently uses the condition of fracture of a unidirectional ber laminate, which is considered as a homogeneous orthotropic material. In this case, one uses generalized strength criteria of the form (5.1.6), see Malmeister et al. (1980), Teters et al. (1978). Then, the functional of restrictions on the strength can be taken in the form
1 (t

); 0 sup j (yy ; y(jy )j 0 y y0 2 0 e4 ] j

c1
o

!(t)) = (z max2T ' )

2 X

2 X

i k=1

() ik (v ( ) ( )) ik (z

' )
(5.7.154)

p q m n=1

( pqmn (v ( ) ( )) pq) (z

() ' ) mn (z ' ) ; 1

T = 0 L]

02 ]

;q q]:

( Here, ik ) is the averaging of the function ik in the z and ' coordinates (see subsec. 5.1.2), where the function ik is determined by (5.7.116) via the function of displacements !(t) ik (v ) and pqmn (v ) are the components of the strength tensor of a unidirectional ber laminate which correspond to the basis rotated at angle with respect to the direction of the bers of the laminate and depend on the volume fraction of bers v. In addition, the functional of restrictions on the strength 2 is also introduced. It is obtained from 1 by substituting ik (v ; ) and pqmn (v ; ) for ik (v ) and pqmn (v ). Suppose that v ! ik (v 0) and v ! pqmn (v 0) are continuous functions on e1 e2 ]. Then, making use of Theorem 5.7.6 and the properties of averaging, one can see that the functional 1 from (5.7.154) satis es the condition (5.7.151).

5.7. Optimization of structures of composite materials

371

As a goal functional, the mass of the shell is often taken. Then, (t) = 4 rL
Zq
0

v( ) f + (1 ; v( )) m ] d :

(5:7:155)

The problem of optimization of the composite laminate shell reduces to the search of an element t0 such that

t0 = (q0

0 v0 ) 2 Uad

Theorem 5.7.7 Let the assumptions of Theorem 5.7.6 be ful lled. Let a nonempty Proof. Let ftng be a minimizing sequence: tn = (qn n vn ) 2 Uad lim (t ) = t2U (t): inf n!1 n ad
ftm g such that qm ! q0 in R
m!
0

(t0 ) = t2U inf

ad

(t):

(5:7:156)

set Uad and a goal functional be de ned by the relations (5.7.151){(5.7.153). Then, the problem (5.7.156) has a solution.

(5:7:157)

(5.7.123), (5.7.124), and (5.7.152) yield that we can choose a subsequence in C ( 0 e4 ])

(5.7.158) and Theorem 5.7.6 imply that

vm ! v0 in C ( 0 e4 ]): (5:7:158)
0

!(tm ) ! !(t0 ) in V

t0 = (q0

v0 ):

(5:7:159)

Now, by virtue of (5.7.151) and (5.7.157){(5.7.159), one can verify that the function t0 = t0 is a solution of the problem (5.7.156). The theorem is proven. By using the above approach, one can examine various problems of optimization of composite laminate shells under restrictions on the strength and stability.

5.7.7 Optimization of the composite structure

In subsec. 5.7.4, we considered the problem of getting the e ective elasticity constants that are most close to the given ones. There, we supposed that the composite had Y -periodic structure, and the shape of the domain G occupied by the inclusion (more exactly, the function f de ning the boundary of G) considered as a control. However, elements of constructions are, as a rule, in inhomogeneous strainstress state. This is why in many cases it is desirable that the e ective elasticity coe cients be some functions of the coordinates of points of the domain occupied by the composite (by an element of a construction made of a composite material). Now, we will suppose that the \period" Y depends on a point y 2 R3 , and the function f de ning the boundary of the inclusion depends not only on a

The set of controls and functions of e ective elasticity coe cients

372

Chapter 5. Optimization of deformable solids

point s of the unit sphere S0 in R3 , but also on a point y 2 . Denote Q = and de ne a set of controls U as follows:

S0 ,

U = r = (Y f ) Y = fYi g3=1 2 (C ( ))3 i (y ; 0 sup jYiky ); yYki (y )j 0 3 R y y0 2 f f f^ sup jf (y

^ Yi Yi Yi

ci

i = 1 2 3 f 2 C (Q) s 2 S0
(5.7.160)

s) ; f (y0 s)j c 4 ky ; y0 kR3 y y0 2 o j ( ;f 0 0 sup 0 kfyyss); (y0(y 0 )s )j6 c5 : s kR (y s) (y0 s )2Q ( )

^ ^ Here, Yi , Yi , i = 1 2 3, and f , f are continuous positive-number-valued functions de ned in and Q, respectively, c1 : : : c5 are given positive constants. For a given element r = (Y = fYi g3=1 f ) from U , with every point y 2 i
e we associate the cube Y (y) =
3 Y

e de ned on the unit sphere S0 . Put also r(y) = (Y (y) f (y )). The inclusion and matrix elasticity constants ain and am are supposed ijkm ijkm e to be given and xed. Points of a cube Y (y) are denoted by x. In accordance with (5.7.78) and (5.7.80), the coe cients ar(y) are de ned on ijkm e the cube Y (y) by

i=1

0 Yi (y)] and the function f (y ): s ! f (y s)

ar(y) (x) = ijkm

where

ain ijkm if x 2 G(r(y )) e am ijkm if x 2 Y (y ) n G(r(y ))

1 s 2 S0

(5.7.161)

G(r(y)) = x j x = x0 (y) + f (y s)n(s) 0

(5.7.162)

and n(s) is the unit vector that is normal at the point s to the surface of the unit 1 sphere S0y centered at the point x0 (y) = f 2 Yi (y)g3=1 . Notice that, on one hand, i we have the xed sphere S0 , which is supposed to be centered at the origin. On the other hand, to every y 2 there corresponds the unit sphere S0y centered at x0 (y). Here, we suppose that the one-to-one correspondence between points of S0 and S0y is determined by the translation of the sphere S0 without rotation. Therefore, we identify points of these spheres and denote them by a single letter s. e Denote by Wr(y) the space W1 de ned in the cube Y (y), see Lemma 5.7.2. Next, according to (5.7.81), on the space Wr(y) we de ne the following bilinear

5.7. Optimization of structures of composite materials

373

form:

ar(y)(u v) =

Z
e Y (y)

3 X

i j k m=1

ar(y) (x)"km (u)"ij (v) dx: ijkm

(5.7.163)

Suppose that the inclusion and matrix elasticity constants satisfy the conditions of symmetry and positive de niteness. Then, upon Lemma 5.7.2, there exists a unique function X ij (r(y)) that is the solution of the following problem:

X ij (r(y)) 2 Wr(y) ar(y) (X ij (r(y)) v) = ar(y) (P ij v)

v 2 Wr(y)

(5.7.164)

where i j = 1 2 3 and P ij are given by (5.7.19). Now, because of (5.7.83), the values of the e ective elasticity coe cients at point y 2 are given by

coe cients

; ar(y) X ij (r(y)) ; P ij X km (r(y)) ; P km gijkm (r(y)) = : (5.7.165) Y1 (y)Y2 (y)Y3 (y) Thus, to every r 2 U there corresponds the functions of e ective elasticity

that are de ned, in the domain occupied by the composite and on its boundary, via the relations (5.7.82) and (5.7.83). In order to justify the use of the latter formulas, we have to suppose that the functions y ! Yi (y) and y ! f (y s) are slowly changing in , i.e., the constants c1 : : : c4 in (5.7.160) are su ciently small.

gijkm (r): y ! gijkm (r(y))

de ned in . In the space H de ned by (5.7.7) and equipped with the norm (5.7.14), we de ne a bilinear form br by

The problem of the stress-strain state of a composite Suppose is a bounded Lipschitz domain in R3 . By S we denote the boundary of . Suppose also that S = S1 S2 , S1 \ S2 = ?, mes S1 > 0, on S1 the composite is fastened, and surface forces act on the part S2 . Thus, for every r 2 U , there are functions of the elasticity coe cients gijkm (r)
br (u v) =
Z
3 X

i j k m=1

gijkm (r)"km (u)"ij (v) dy

u v 2 H:

(5.7.166)

Let the composite be a ected by a xed load that does not depend on control. This load is identi ed with an element L 2 H . Consider the problem: For a xed r 2 U , nd a function wr such that

wr 2 H

br (wr v) = (L v)

v 2 H:

(5.7.167)

374

Chapter 5. Optimization of deformable solids

Theorem 5.7.8 Let a set U be de ned by (5.7.160), L 2 H , and let a bilinear form br be de ned by the formula (5.7.166) in which the functions y ! gijkm (r(y))

are given by the relations (5.7.161){(5.7.165). Let also ain , am be constants ijkm ijkm satisfying the conditions of symmetry and positive de niteness (5.7.4), (5.7.5). Then, there exists a unique solution of the problem (5.7.167).

Proof. From Remark 5.7.3, it follows that, at every point y 2 , the coe cients gijkm (r(y)) satisfy the condition of symmetry, and the following estimate holds:
3 X

i j k m=1

gijkm (r(y))akm aij

r2U y2 :

3 X 2 c0 aij i j =1

aij = aji 2 R

(5.7.168)

From the proof of Theorem 5.7.2 (see (5.7.48){(5.7.50)), we conclude the existence of a constant c such that
y2

sup jgijkm (r(y))j c

r2U

(5.7.169)

for arbitrary indices i j k m = 1 2 3. So, for any r 2 U , the bilinear form br is symmetric, continuous, and coercive in H . Hence, there exists a unique solution of the problem (5.7.167). In virtue of Theorem 5.7.8, we can de ne the function r ! wr mapping the set U in the space H . The theorem below establishes the continuity of this function.

Theorem 5.7.9 Under the conditions of Theorem 5.7.8, the function r ! wr is

a continuous mapping of U (in the topology generated by the product of (C ( ))3 and C (Q)) in the space H:

To prove Theorem 5.7.9, we need the following lemma. Lemma 5.7.6 Let the conditions of Theorem 5.7.8 be satis ed, let

frn = (Yn fn)g1=1 n be a sequence of elements of U , and let Yn ! Y0 in (C ( ))3 , fn ! f0 in C (Q). Then, r0 = (Y0 f0) 2 U and (5.7.170) lim sup gijkm (rn (y)) ; gijkm (r0 (y)) = 0 n!1
y2

for arbitrary indices i j k m = 1 2 3: Proof. By using (5.7.160), it is not hard to see that the conditions

frn = (Yn fn )g1=1 U Yn ! Y0 in (C ( ))3 fn ! f0 in C (Q) n

5.7. Optimization of structures of composite materials

375

imply that r0 = (Y0 f0 ) 2 U . Let us prove the equality (5.7.170). Assume rst that Yn = Y0 for each n. Since fn ! f0 in C (Q), from the proof of Lemma 5.7.4, it follows that rn (y) ; ar0 (y) k (5.7.171) L2 (Wr0 (y) Wr0 (y) R) = 0: nlim sup ka !1
y2

Denote by A(rn (y)) the operator generated by the bilinear form arn(y) , n = 0 1 2 : : : , and given by the formula A(rn (y)) 2 L(Wr0 (y) Wr0 (y) ) (A(rn (y))u v) = arn (y)(u v) u v 2 Wr0 (y) : (5.7.172) From (5.7.171), we infer that lim sup kA(rn (y)) ; A(r0 (y))k = 0: (5.7.173) n!1
y2

As well known (see, e.g., Schwartz (1967)), if p is a linear, continuous, invertible mapping of a Banach space E into a Banach space F and if q is a linear continuous mapping of E into F satisfying the inequality kqk < kp;1k;1 , then the mapping p + q is invertible, and the following expansion holds: (p + q);1 = p;1 ; p;1 qp;1 + p;1 qp;1 qp;1 ; : (5.7.174) Thus, letting p = A(r0 (y)), q = A(rn (y)) ; A(r0 (y)) and taking (5.7.173) into account, we have that, for su ciently large n, say n K , k(A(rn (y)));1 ; (A(r0 (y)));1 k (5.7.175) c1 kA(rn (y)) ; A(r0 (y))k k(A(r0 (y)));1 k2: From the positive de niteness of the coe cients ain and am , we conclude ijkm ijkm the existence of a constant c2 such that
y2

sup A(r0 (y))

;1

c2 :
(5.7.176)

Now, from (5.7.173) and (5.7.175), we get ;1 ;1 nlim sup k(A(rn (y ))) ; (A(r0 (y ))) k = 0: !1
y2

By using (5.7.164), (5.7.165), (5.7.171), and (5.7.176), it is not hard to derive (5.7.170). Assume now that the condition Yn = Y0 for all n does not hold. Then, for an arbitrary xed point y 2 , to every n there corresponds the cube
e Yn (y) =
3 Y

i=1

0 Yin (y)]

376

Chapter 5. Optimization of deformable solids

and the bilinear forms arn (y) are de ned on functions given on di erent cubes e Yn (y). For a xed y, introduce for each index n the variables xn = (xn xn xn ) that 1 2 3 are connected with the variable x = (x1 x2 x3 ) by the equality
i ( xn = xYYi0 yy) : i ()

In the xn variables, the problem of nding the elasticity coe cients at any point Q y 2 with di erent n reduces to the problem on the xed domain 3=1 (0 Yi0 (y)), i and analogously to the above, we establish (5.7.170).

in

Remark 5.7.8 Analogously to the proof of Lemma 5.7.6, one can show that, for an arbitrary xed r 2 U , the function y ! gijkm (r(y)) is continuous in , and therefore, in (5.7.170), the supremum can be replaced with the maximum.
the condition (5.7.170) is satis ed. Hence, (5.7.166) implies

Proof of Theorem 5.7.9. Let frn = (Yn fn)g1=1 be a sequence of elements of U , n r0 = (Y0 f0 ) 2 U , Yn ! Y0 in (C ( ))3 , fn ! f0 in C (Q). Then, by Lemma 5.7.6,
brn ! br0 in L2 (H H R):
Denote by Br the operator generated by the bilinear form br : (Br u v) = br (u v) (5.7.177)

u v 2 H:

By (5.7.177), we get Brn ! Br0 in L(H H ). Upon (5.7.168), the operators Brn and Br0 are invertible. So, Theorem 1.8.1 gives wrn ! wr0 in H , concluding the proof.

Optimization problem

Suppose that we are given functionals k such that 9 k : r w ! k (r w) is a continuous mapping of U H > = (in the topology generated by the product of the topologies> (C ( ))3 , C (Q), and H ) into R, k = 0 1 : : : m. De ne a set of admissible controls Uad as follows:

(5.7.178)

Uad = r j r 2 U

k (r wr )

0 k = 1 2 ::: m

(5.7.179)

where wr is a solution of the problem (5.7.167). Let also a goal functional be of the form (r ) =
0 (r

wr ):

(5.7.180)

5.8. Optimization of laminate composite covers

:::

377

Consider the problem of nding r such that ~

r 2 Uad ~

(~) = r2U r inf

ad

(r):

(5.7.181)

By using Theorem 5.7.9, analogously to the above (see, for example, the proof of Theorem 5.7.3), one establishes the following
and a goal functional be de ned by (5.7.178){(5.7.180). Then, there exists a solution of the problem (5.7.181).

Theorem 5.7.10 Under the conditions of Theorem 5.7.8, let a nonempty set Uad

zr : ! R as follows:

We will present now some examples of k and . For r 2 U , de ne a function

zr (y) = Y (y)Y 1y)Y (y) dx 1 2( 3 Gr (y)

where Gr (y) is de ned by (5.7.162). The goal functional can be taken in the form (r ) =
Z

(5.7.182)

zr (y) dy:

(5.7.183)

The formula (5.7.182) for a given control r determines the ratio of the volume e of the inclusion at a point y to the volume of the cube Y (y). Hence, the choice of the goal functional in the form (5.7.183) means the minimization of the fraction of the inclusion in the volume of the composite. Usually, the speci c gravity and cost of the inclusion are greater than those of the matrix, so that the minimization leads to decrease of the weight of the composite, as well as to decrease of its cost. If, nevertheless, the speci c gravity and cost of the matrix are greater than those of the inclusion, one can choose the goal functional in the form (r ) = ;
Z

zr (y) dy:

The functionals k , k = 1 2 : : : m, can correspond to the restrictions on sti ness and strength.

Remark 5.7.9 There are approaches to the optimal design of composites and materials of structures that are based on the creation of a speci c mixture of layers at each point of the structure. These approaches have been initiated by Olho (1974), Lurie and Cherkaev (1976), Armand and Lodier (1978). For further results, see Rozvany (1989), Lurie (1993), Bends e (1994).

378

Chapter 5. Optimization of deformable solids

5.8 Optimization of laminate composite covers according to mechanical and radio engineering characteristics
In radio-location, one faces di erent problems of design of the protective devices of antenna systems (radar domes), which are usually made of laminate composite materials and must satisfy some thermal, mechanical, and radio engineering demands more exactly, they must be strong, sti , and radioparent in a su ciently large wave band or in some scattered wave bands, besides they should have a small weight. Sometimes, one also faces problems of design of absorbers of radio waves, which are also made of laminate composite materials. These should be strong and sti , and should maximally absorb radio waves, having a small weight. Below, in subsec. 5.8.1, we will consider the problem of propagation of radio waves through a laminate medium. In subsec. 5.8.2, we will consider some optimization problems according to mechanical and radio engineering characteristics.

We will be concerned now with a method of calculation of characteristics of the propagation of plane harmonic waves through a laminated medium. Our presentation is based on the classical work by Abeles (1950), see also Born and Wolf (1964).

5.8.1 Propagation of electromagnetic waves through a laminated medium

Basic equations
The Maxwell equations have the following form

" @E = rot H c @t @H c @t = ; rot E:

(5.8.1) (5.8.2)

Here, E = (Ex Ey Ez ) is the vector of electric eld, H = (Hx Hy Hz ) the vector of magnetic eld, t the time, c the velocity of light in the vacuum, " the dielectric permittivity, the magnetic permeability. We consider a laminated medium such that " and are functions of z only. Moreover, is a real function, and " is a real function for a nonconducting medium, and a complex function for a conducting medium in the latter case, " is said to be the complex dielectric permittivity. We suppose the electromagnetic wave to be plane. Let yOz be a plane of incidence. Let TE stand for a transverse electric wave such that the electric vector is perpendicular to the yOz plane, i.e., E = (Ex 0 0). Let also TM stand for a transverse magnetic wave for which the magnetic vector is perpendicular to the

5.8. Optimization of laminate composite covers

:::

379

yOz plane, i.e., H = (Hx 0 0). An arbitrary plane wave may be represented as a sum of TE and TM waves.

Consider a TE wave. We seek the solution of the equations (5.8.1), (5.8.2) in the form e e e e e E (x y z t) = E (x y z ) exp(;i!t) E = (Ex Ey Ez ) (5.8.3) e (x y z ) exp(;i!t) e = (Hx Hy Hz ) e e e H (x y z t) = H H (5.8.4) where ! is the angular frequency. Then, the vector equations (5.8.1), (5.8.2) reduce to the following scalar equations:
e e @ Hz ; @ Hy + i"! E = 0 e @y @z c x e e @ Hx ; @ Hz = 0 @z @x e y @ Hx @H ; e = 0 @x @y i! H = 0 e c x e @ Ex ; i! H = 0 e @z c y e @ Ex + i! H = 0: e @y c z

(5.8.5a) (5.8.5b) (5.8.5c) (5.8.6a) (5.8.6b) (5.8.6c)

In case of a conducting medium, the complex dielectric permittivity is given by (see Born and Wolf (1964)) " = "0 + i 4 (5.8.6d) where " is the dielectric permittivity, the conductance. e e e e The equations (5.8.5a){(5.8.6c) show that Hx = 0 and Ex , Hy , Hz depend on y, z only. By (5.8.5a), (5.8.6b), and (5.8.6c), we get
0

e e e @ 2Ex + @ 2 Ex + " k2 E = d(log ) @ Ex ex 2 2 @y @z dz @z

(5.8.7) (5.8.8)

where

We seek the solution of (5.8.7) in the form of a product of two functions, one depending on y and the other one on z :
e Ex (y z ) = Y (y)U (z ):

k = !: c

(5.8.9)

380 Then, (5.8.7) gives

Chapter 5. Optimization of deformable solids

1 d2 Y = ; 1 d2 U ; " k2 + d(log ) 1 dU : (5.8.10) Y dy2 U dz 2 dz U dz The left hand side of this equation depends on y only, the right hand side only on z , therefore

d2 U ; d(log ) dU + " k2 U = k2 2 U dz 2 dz dz
where have

1 d2 Y = ;k2 Y dy2

(5.8.11) (5.8.12)

is a constant (possibly, complex). By (5.8.3), (5.8.9), and (5.8.11), we

Ex = U (z ) exp i(k y ; !t) : (5.8.13) It follows from (5.8.5 a), (5.8.6 b), (5.8.6 c) that Hy and Hz are given by the
analogous formulas

Hy = V (z ) exp i(k y ; !t) Hz = W (z ) exp i(k y ; !t)

with

(5.8.14) (5.8.15) (5.8.16) (5.8.17) (5.8.18) (5.8.19)

dV = ik( W + "U ) dz dU = ik V dz U + W = 0: dU = ik V dz dV = ik " ; dz

By substituting W from (5.8.18) into (5.8.16), we obtain the following system

2

U:

(5.8.20)

The system (5.8.1), (5.8.2) is symmetric in the sense that, if one interchanges the positions of E and H and replaces " by ; and by ;", then the system will remain invariable. Therefore, by using such interchange and replace, any result for a TM wave may be derived from a corresponding result for a TE wave, Thus, for a TM wave, we get Ex = 0 and

Hx = U (z ) exp i(k y ; !t)

(5.8.21)

5.8. Optimization of laminate composite covers

:::

381 (5.8.22) (5.8.23) (5.8.24) (5.8.25) (5.8.26)

Ey = ;V (z ) exp i(k y ; !t) Ez = ;W (z ) exp i(k y ; !t)

and moreover,

dU = ik"V dz dV = ik ; 2 U dz " U + "W = 0:

We consider a stack of s homogeneous slabs, see Fig. 5.8.1. The rst slab occupies the space from z = 0 to z = z1 , the second one from z = z1 to z = z2 , and so forth. In this case, "(z ) = "n z 2 (zn;1 zn) (5.8.27) (z ) = n z 2 (zn;1 zn ) n = 1 : : : s z0 = 0 where "n , n are constants.
y

Characteristic matrix, re ectivity, and transmissivity

j j

0 0

j z1 z2 zs1

s+ 1

zs

Figure 5.8.1: Stack of homogeneous slabs We suppose that a homogeneous medium occupies the space z < 0 "0 and are the dielectric and magnetic constants of this medium. The x component of the electric eld in the medium is given by Ex = E0 exp ikp"0 0 (y sin '0 + z cos '0 ) ; i!t z<0 where '0 is the angle of incidence. From the continuity of the tangential component of the vector E at z = 0, we obtain the following relation for a TE wave: E0 exp i(kp"0 0 y sin '0 ; !t) = U (0) exp i(k y ; !t) 8y: (5.8.28)
0

382 Thus,

Chapter 5. Optimization of deformable solids

= p"0

sin '0 :

(5.8.29)

From the continuity of the tangential component of the vector H at z = 0, we nd that (5.8.29) holds also for a TM wave. Now, let

Un = U (zn )

Vn = V (zn )

n = 0 1 : : : s ; 1 z0 = 0:

(5.8.30)

Taking (5.8.27) into account, it is not hard to verify that, for a TE wave, the solution U , V of the problem (5.8.19), (5.8.20) with the initial data Un , Vn is de ned by

U (z ) = Un Fn (z ) + Vn fn (z ) V (z ) = Un Gn (z ) + Vn gn(z )
Here,

z 2 zn zn+1 ]:

(5.8.31)

Fn (z ) = cos kn (z ; zn ) Gn (z ) = ipn+1 sin kn (z ; zn )

where
r

fn (z ) = p i sin kn (z ; zn ) n+1 gn (z ) = cos kn (z ; zn ) (5.8.32)

pn+1 = "n+1 cos 'n+1

n+1

kn = kp"n+1 n+1 cos 'n+1

(5.8.33)

and
2 0 cos2 'n+1 = 1 ; "" 0 sin '0 :

n+1 n+1

(5.8.34)

sin '0 = p"n+1 n+1 sin 'n+1 n = 0 1 ::: s (5.8.35) and if "0 , "n+1 are real constants, then 'n+1 is the angle between the normal to the plane of constant phase and the Oz axis. The formulas (5.8.31){(5.8.33) are valid for a TE wave, but they also hold true for a TM wave under the condition
0 0

get

The formula (5.8.34) follows from the condition = const, and moreover, we

p"

pn+1 = " n+1 cos 'n+1 n+1

(5.8.36)

5.8. Optimization of laminate composite covers

:::

383

Hx , Ey , Ez de ned by (5.8.21){(5.8.23), (5.8.29). The formula (5.8.34) holds also

for a TM wave. De ne matrices Nn (z ) by

F (z ) Nn (z ) = Gn (z ) fn(z ) (z ) gn n
By (5.8.31), (5.8.37), we have

z 2 zn zn+1]:

(5.8.37) (5.8.38)

U (z ) = N (z ) Un : n V (z ) Vn The matrix Nn (z ) is invertible as det Nn (z ) = 1 for an arbitrary z 2 denote its inverse by Mn(z ), g f Mn (z ) = ;Gn (z ) ;Fn (z ) : n (z ) n (z )
Then

zn zn+1 ]
(5.8.39) (5.8.40)

Un = M (z ) U (z ) n Vn V (z )

z 2 zn :zn+1] n = 0 1 : : : s ; 1

and we have

U0 = M (z )M (z ) M (z )M (z ) U (z ) 0 1 1 2 n;1 n n V0 V (z ) z 2 zn zn+1 ]: Here, U0 , V0 are the values of U , V at z = 0 (see (5.8.30)).

(5.8.41)

The re ection and transmission factors

Introduce the matrix

M = M0 (z1 )M1 (z2 )

Ms;1 (zs )

M = m11 m12 : m21 m22

(5.8.42)

Let A, R, T be the amplitudes (possibly, complex) of the electric vectors of incident, re ected, and refracted (passed) waves, respectively. The boundary condition is the continuity of the tangential components of the vectors E and H at z = 0, see Fig. 5.8.1. Taking into account that

H= "e E

(5.8.43)

where e is the unit vector of wave directivity, and the sign stands for the vector product, we obtain the following relations for a TE wave: U0 = U (0) = A + R U (zs ) = T

384

Chapter 5. Optimization of deformable solids

V 0 = p 0 (A ; R )
where
r p0 = "0 cos '0
0

V (zs ) = ps+1 T ps+1 =

r

(5.8.44) (5.8.45)

"s+1 cos ' : s+1

s+1

Here, "s+1 , s+1 are the dielectric permittivity and the magnetic permeability of a homogeneous medium occupying the half-space z > zs , cos 's+1 de ned by (5.8.34) at n = s. By applying (5.8.41), (5.8.42), we obtain from (5.8.44) that

A + R = (m11 + m12 ps+1 )T p0 (A ; R) = (m21 + m22 ps+1 )T:

(5.8.46)

These equations lead us to the following expressions for the re ection and transmission factors: (5.8.47) r = R = (m11 + m12 ps+1 )p0 ; (m21 + m22 ps+1 ) A (m11 + m12 ps+1 )p0 + (m21 + m22 ps+1 ) T 0 t = A = (m + m p )p2p+ (m + m p ) : (5.8.48)
11 12 s+1 0 21 22 s+1

The re ection and transmission powers are de ned by

R = jrj2
r

F = ps+1 jtj2 : p
0

(5.8.49)

The formulas (5.8.47){(5.8.49) were given for a TE wave, but these also valid for a TM wave under the conditions

p0 =

0 "0

cos '0

ps+1 =

"s+1 cos 's+1 :

s+1

(5.8.50)

In this case, r and t are the ratios of the amplitudes of magnetic vectors, but r is also the ratio of the amplitudes of electric vectors. For a TM wave, we denote by t1 the ratio of the amplitudes of the tangential components of electric vectors of transmitted and incident waves. Then, by (5.8.43), we get r s t1 = s+1 "0 cos ''+1 t (5.8.51) 0 "s+1 cos 0

t de ned by (5.8.48), (5.8.50). The transmission power for a TM wave may be

determined by
s F = pp+1 jt1 j2 :
0

5.8. Optimization of laminate composite covers

:::

385

5.8.2 Optimization problems

We consider a cover (a radar dome, or an absorber) as a shell made of laminated composite materials. The stress-strain state of the laminated shell is de ned by a function of displacements ! which is the solution of the following problem ~

!2V ~ a (~ !) = (f !) ! ! 2 V: (5.8.52) Here, V is a subspace of W that satis es (4.6.11), W de ned by (4.6.8) and (4.7.5) for a shell of revolution and for a shallow shell, respectively, the bilinear form a for these shells determined in subsecs. 4.12.2 and 4.12.3, f is the load, f 2 V , The cover is subject to various loads. Denote by T the set of the loads on
which the cover is calculated, and suppose that T is a compact set in V . Introduce the set B= =( 1
2

(5.8.53)

: : : s ) 2 Rs j

a1 n ; n;1 a1 n = 2 3 : : : s s a2 a1 , a2 are positive constants :

1

(5.8.54)

By Theorems 4.12.1 and 4.12.2, for each 2 B and f 2 T for the shell of revolution and for the shallow shell, there exists a unique solution ! of the problem ~ (5.8.52) denote this solution by !( f ). So, ~

!( f ) 2 V ~

a (~ ( f ) !) = (f !) !

! 2 V:

(5.8.55)

B is considered as a set of controls, i.e., the number of laminae is supposed n to be xed and equal to s. The mechanical parameters Eij (see (4.12.6)) and the electromagnetic parameters "n , n (see (5.8.27)) are also supposed to be xed. For each 2 B , the state of the cover Q is a set of functions !( f ), where f runs ~ through T , i.e., Q = !( f ) f 2T : ~
Suppose that Ai : ! ! Ai ( !) is a continuous mapping of B V into R, i = 1 2 : : : q. De ne mappings (5.8.56) (5.8.57) (5.8.58)

B V 3 ( f ) ! ;i ( f ) = Ai ( !( f )) 2 R ~

i = 1 2 : : : q:

The functionals Ai may be determined in such a way that ;i ( f ) q=1 de ne i stresses, displacements, and some functions of them.

386

Chapter 5. Optimization of deformable solids

We determine functionals i on B by ! i ( ) = sup ;i ( f )

f 2T q+1 ( ) = s X i=1

i = 1 2 ::: q bi i :

(5.8.59) (5.8.60)

Here, bi are positive constants that may be associated with the speci c gravity or cost of the materials of laminae. We de ne a set of admissible controls as follows: Uad = 2 B j i ( ) ci i = 1 : : : q + 1 ~ (5.8.61) where ci are positive constants. The relations i ( ) ci can be constructed in the ~ ~ form of restrictions on strength, sti ness, weight, cost, and so forth. Consider some goal functionals. Suppose that the modules of the radii of the principal curvatures of the shell are 1:5 2 times greater than the wavelength = 2 c :
0

!p"0

This condition holds usually in practice. Then, we can apply the formulas (5.8.47){ (5.8.49). We suppose that the surface = h is the surface of incidence of the wave. 2 So, in (5.8.42), we should take (see Figs. 4.12.1 and 5.8.1) zn = h ; s;n n = 1 2 ::: s (5.8.62) 0 = 0: The re ection and transmission factors de ned by (5.8.47), (5.8.48) depend on the vector = ( 1 2 : : : s ) 2 B , on the angle of incidence of the wave '0 , and on the wavenumber k from (5.8.8). Thus, we denote the re ection and transmission factors for TE and TM waves by rE ( '0 k), tE ( '0 k), rM ( '0 k), tM ( '0 k) for these waves p0 and ps+1 are de ned by (5.8.45), (5.8.50). In case of a radar dome, we should maximize the radioparency. So, we can take the following goal functional
0( ) = ;

n1 n2 XX i=1 j =1

aij jtE ( '0i kj )j2 + bij jtM ( '0i kj )j2

(5.8.63)

which must be minimized. Here, '0i , kj are given values of the angles of incidence and the wave numbers, n1 , n2 nite numbers, aij , bij weight factors, positive constants. In case of an absorber, we should minimize the re ection and transmission powers, i.e., the following functional
n1 n2 X X n (1) aij jrE ( 0( ) = i=1 j =1

'0i kj )j2 + a(2) jrM ( '0i kj )j2 ij

5.9. Shape optimization of two-dimensional elastic body

387
o

+ a(3) jtE ( '0i kj )j2 + a(4) jtM ( '0i kj )j2 ij ij

(5.8.64)

where a(k) are positive constants. ij Consider the following optimization problem: Find a vector = ( 1 2 : : : , s ) satisfying 2 Uad inf 0 ( ) (5.8.65) 0 ( ) = 2U
ad

where 0 is given by (5.8.63) or (5.8.64). We consider the cases when the cover is either a shell of revolution or a shallow shell. Theorem 5.8.1 Suppose either the conditions of Theorem 4.12.1 (for a shell of revolution ) or of Theorem 4.12.2 (for a shallow shell ) are satis ed. Let functionals i and a nonempty set Uad be de ned by (5.8.53), (5.8.54), (5.8.57){(5.8.61). Suppose a goal functional is determined by (5.8.63) or (5.8.64). Then, there exists a solution of the problem (5.8.65). Proof. Obviously, B is a compact set in Rs . Therefore, there exists a sequence f n g1=1 such that n n 2 Uad 8n (5.8.66) lim 0 ( n ) = 2U 0 ( ) inf (5.8.67) Let A be the operator generated by the form a , i.e., A 2 L(V V ) (A u v) = a (u v) u v 2 V: (5.8.69) By (5.8.68), (4.12.6), and (4.12.8), we get A n ! A 0 in L(V V ). Applying Theorem 1.8.1, we obtain ! : f ! !( f ) is a continuous mapping of B V into V , ~ ~ (5.8.70) where !( f ) is the solution of the problem (5.8.55). Theorem 1.4.4 and (5.8.53), ~ (5.8.57){(5.8.59), (5.8.70) give that i 's are continuous functionals in B , i = 1 : : : q. Therefore, 0 2 Uad . Due to (5.8.32), (5.8.39), (5.8.42), (5.8.47), and (5.8.48), we conclude that 0 is a continuous functional on B . So, the vector = 0 is a solution of the problem (5.8.65).
n ! 0 in Rs
ad

0 2 B:

(5.8.68)

5.9 Shape optimization of two-dimensional elastic body

Earlier, we have studied optimization problems for deformable solids in which the weak (generalized) solutions of the state equations in the form of elliptic equations

388

Chapter 5. Optimization of deformable solids

were used. However, in some formulations of optimization problems for deformable solids with restrictions on strength, the stresses must be bounded at each point of a deformable solid. So, in such a case, smooth solutions of the state equations are only allowed. In this and next sections, we will consider two problems of such type for elastic bodies, using results of Section 2.13.

5.9.1 Sets of controls and domains in the optimization problem

Let M be a space of controls, which will be de ned below. A domain q in R2 occupied by an elastic body is given for every q 2 M . The boundary Sq of q consists of two connected components S1 and S2q , see Fig. 5.9.1. The points of S1 are held xed, and S1 does not depend on a control q. Surface forces F = (F1 F2 ) are given on S21 , which is an open set in S2q , and these forces are continued onto the whole S2q by zero. We suppose also that S21 does not depend on a control q, (1) (1) and S2q = S2q n S21 is the controlled part of S2q , that is, S2q is the part of the boundary that must be chosen from the optimization conditions.
y2
q(a )

F = ( F1 , F2 )

S1

y1

S2 q

Figure 5.9.1: Domain q occupied with an elastic body De ne the space of controls in the form
e M = q j q 2 C l]+3 ( 0 2 ]) r1 < q( ) < r2 q( ) = ( ) 2 ( 1 2 ) :

202 ]

(5.9.1)

Here, r1 , r2 are positive constants, r1 < r2 , l > 0, l not an integer, l] is an integer such that l ; l] 2 (0 1), is a function de ned on ( 1 2 ) (0 2 ), and in polar coordinates r, the part S21 of S2q consists of points s = (r ) such that 2 e ( 1 2 ) and r = ( ). C l]+3 ( 0 2 ]) is the subspace of C l]+3 ( 0 2 ]) consisting of the periodical functions. The periodicity of a function q 2 C l]+3 ( 0 2 ]) means that, if q is the 0 2 ]-periodical continuation of q on R, then q 2 C l]+3 ( a b]) for ~ ~

5.9. Shape optimization of two-dimensional elastic body

389

an arbitrary a b] R. The set M is equipped with the topology generated by that of C l]+3 ( 0 2 ]). For each q 2 M , de ne a domain q such that the internal boundary S1 of e it is given in polar coordinates by a function 2 C l]+3 ( 0 2 ]), and the external boundary S2q by the function q. De ne the domain in the form = x j x = (x1 x2 ) 1 < x2 + x2 < 4 : (5.9.2) 1 2 Let E stand for the function that maps polar coordinates into Cartesian ones: E : (r ) ! E (r ) = (y1 y2 ) y1 = r cos y2 = r sin (5.9.3) and let E ;1 be its inverse. Determine Pq : q ! by the formula Pq = E Gq E ;1 where Gq : E ;1 ( q ) ! E ;1 ( ), (r ) ! Gq (r ) = ( ') ( = r ; (2 )( ;) + q) ) '= : q ( The mapping G;1 : E ;1 ( ) ! E ;1 ( q ) has the form q (5.9.4)

(5.9.5)

( ') ! G;1 ( ') = (r ) q (5.9.6) r = 2 (') ; q(') + q(') ; (')] = ': It is easily seen that the mapping Pq de ned by (5.9.4), (5.9.5) is a di eomorphism of q onto of the C l]+3 class.

5.9.2 Problems of elasticity in domains

Aq u = >
> : 8 > > <

In the domain q , the operator Aq of the theory of elasticity has the form

Here, u = (u1 u2) is a vector function of displacements, , are positive constants. Denote by "ij (u), ij (u) the components of the strain and stress tensors: 1 @u "ij (u) = 2 @y i + @uj j @yi (5.9.8) @u1 + @u2 i j=1 2 ij (u) = @y @y ij + 2 "ij (u)
1 2

@ 2 u1 + @ 2 u1 + ( + ) @ 2 u1 + @ 2 u2 2 2 2 @y1 @y2 @y1 @y1@y2 @ 2 u2 + @ 2 u2 + ( + ) @ 2 u1 + @ 2 u2 2 2 2 @y1 @y2 @y1@y2 @y2

9 > > = > >

(5.9.7)

390

Chapter 5. Optimization of deformable solids

ij the Kroneker delta. De ne boundary operators B and Bq on S and S2q , respectively, by the expressions

u Bu = u1 S1 2 S1

( Bq u = (

11 (u) 1q + 12 (u) 2q ) S2q 21 (u) 1q + 22 (u) 2q ) S2q

(5.9.9)

where iq are the components of the unit outward normal to S2q , i = 1 2. Theorem 5.9.1 Let a set M be de ned by (5.9.1). For each q 2 M , determine a two-connected domain q R2 such that the internal and external boundaries of e q are de ned in polar coordinates by the functions 2 C l]+3 ( 0 2 ]) and q . Then, the operator Lq : u ! Lq u = (Aq u Bu Bq u) de ned by (5.9.7), (5.9.9), where , are positive constants, is an isomorphism of the space Vlq = C l+2 ( q )2 onto the space Hlq = C l ( q )2 C l+2 (S1 )2 C l+1 (S2q )2 :

Proof. Consider the problem

Aq u = f Bu = g1 Bq u = g2
in q on S1 on S2q (5.9.10)

where (f g1 g2 ) 2 Hlq . The ellipticity of the operator Aq is implied by the Korn inequality. The ellipticity of the problem (5.9.10) follows from the ellipticity of the rst and second problems of the theory of ellipticity (see Michlin (1973)). The kernel space of the operator Aq is the space of small rigid displacements, which has the form (see subsec. 1.7.2)

Q = u j u = (u1 u2) u1 = a1 + a3 y2 u2 = a2 ; a3 y1 a1 a2 a3 2 R :

(5.9.11)

(1) (1) (2) (2) Let y(1) = (y1 y2 ), y(2) = (y1 y2 ) be two di erent points of S1 . From (1) ) = u(y (2) ) = 0, and if additionally u 2 the condition Bu = 0, we have u(y Q, then by (5.9.11) we get a1 = a2 = a3 = 0. Therefore, the kernel space of the operator Lq = (Aq B Bq ) consists only of zero. For each (f g1 g2) 2 Hlq , there exists a solution of the problem (5.9.10). Hence, the theorem follows from Theorem 2.13.1.

5.9.3 The optimization problem

For each q 2 M , consider the problem

Aq uq = 0 Buq = 0 Bq uq = F

in q on S1 on S2q

(5.9.12)

5.9. Shape optimization of two-dimensional elastic body

391

where the operators Aq , B , Bq are de ned by (5.9.7), (5.9.9). We suppose that

F 2 C l+1 (S2q )2
where (see (5.9.1) and (5.9.3))

supp F

S21 S2q

(5.9.13)

S21 = s j s = E ( ( ) )

2(

2)

Now, we will consider restrictions on strength. For a vector function of displacements u = (u1 u2), the components of the stress deviator (shear stress tensor) are de ned by the formula 1 i j=1 2 ij (u) = ij (u) ; 2 ( 11 (u) + 22 (u)) ij and the second invariant of the stress deviator has the form

T (u) =

2 X

i j =1

( ij (u))2 = 1 ( 2

11 (u) ; 22 (u))

2 + 2(

12 (u))

2:

(5.9.14)

De ne a functional G1 over M by

G1 (q) = max (T (uq ))(y) ; b

y2 q

(5.9.15)

where uq is the solution of the problem (5.9.12), b a positive constant. For an isotropic material, a restriction on strength may be taken in the form G1 (q) 0. The volume of the material is de ned by

G0 (q) =
De ne a set M1 by

dy: c1

(5.9.16)

e M1 = q j q 2 M q 2 C l+3 ( 0 2 ]) kqkC l+3( 0 2 e r1 + q( ) r2 ; 202 ]

])

(5.9.17)

where M is determined by (5.9.1), c1 , are positive constants, and is small. We take a set of admissible controls Uad in the form

Uad = q j q 2 M1 G1 (q) 0 :
The optimization problem consists in nding q0 satisfying

(5.9.18) (5.9.19)

q0 2 Uad

G0 (q0 ) = q2U G0 (q): inf

ad

392

Chapter 5. Optimization of deformable solids

Theorem 5.9.2 Let operators Aq , B, Bq be de ned by the formulas (5.9.7), (5.9.9), and let , be positive constants. Suppose the condition (5.9.13) holds, and functionals G1 and G0 over M are de ned by (5.9.15), (5.9.16), where uq is the solution of the problem (5.9.12). Let also a nonempty set Uad be given by the expressions (5.9.1), (5.9.17), (5.9.18). Then, for any l > 0, l not an integer, there exists a solution of the problem (5.9.19). Proof. De ne spaces Vl and Hl by
Vl = C l+2 ( )2 Hl = C l ( )2 C l+2 (S01 )2 C l+1 (S02 )2 :
Here, is the domain de ned by (5.9.2), S01 and S02 are the internal and external boundaries of . Just as in subsec. 2.13.2, the problem (5.9.12) reduces to the following one: Find a function uq 2 Vl satisfying ~
e ~ Aq uq = 0 e~ B uq = 0 e ~ Bq uq = F Pq;1 e e e Here, the operators Aq , B , Bq are de ned by e Aq u = (Aq (u Pq )) Pq;1 e (5.9.21) Bu = (B (u Pq )) Pq;1 eq u = (Bq (u Pq )) Pq;1 : B From (5.9.12), (5.9.20), and (5.9.21), it follows that uq = uq Pq . It is easily seen ~

in on S01 on S02 :

(5.9.20)

that

e e e e q ! Lq = (Aq B Bq ) is a continuous mapping of M into L(Vl Hl ).

(5.9.22)

Upon (5.9.13), q ! F Pq;1 is a constant mapping, and from Theorem 2.13.3 we now obtain that

q ! uq is a continuous mapping of M into Vl , ~

(5.9.23)

where uq is the solution of the problem (5.9.20). Under the change of variables ~ corresponding to the mapping Pq , the functionals G1 and G0 from (5.9.15), (5.9.16) take the form

G1 (q) = max(T (~q Pq ))(Pq;1 (x)) ; b u

x2

(5.9.24)

5.10. Optimization of the internal boundary

:::

393 (5.9.25)

G0 (q) =

det (Pq;1 )0 (x) dx:

Here, (T (~q Pq ))(Pq;1 (x)) is the value of the function T (~q Pq ) at a point Pq;1 (x), u u and (Pq;1 )0 (x) is the value of the Frechet derivative of the mapping Pq;1 at a point x. Taking into account (5.9.23){(5.9.25), (5.9.4){(5.9.6), and Theorem 1.4.4, we infer that G1 and G0 are continuous functionals over M . As the imbedding of C l+3 ( 0 2 ]) into C l]+3 ( 0 2 ]) is compact, we have that M1 is a compact set in M . Therefore, there exists a solution of the problem (5.9.19).
e e e Remark 5.9.1 In the considered case, the function q ! (Aq B Bq ) is a continuously Frechet di erentiable mapping of M into L(Vl Hl ). Thus, by using Theorem 2.13.4, we obtain that q ! (q) = uq is a continuously Frechet di erentiable ~ mapping of M into Vl . However, the functional G1 from (5.9.24) is not di erentiable, because so is not the functional v ! max v(x), v 2 C ( ). Nevertheless, x2 on the set Q = v j v 2 C ( ) v(x) 0 x 2 , the latter functional may be approximated by the continuously Frechet di erentiable functional v ! kvkLp( )

if p is su ciently large.

5.10 Optimization of the internal boundary of a two-dimensional elastic body

In Section 5.9, we considered the optimization problem for a two-connected elastic body in which we assigned displacements on the internal boundary and surface forces on the external boundary. In that case, there exists a unique solution of the problem (5.9.10) for any (f g1 g2 ) 2 Hlq , and owing to Theorem 5.9.1 the condition (2.13.9) is satis ed. Now, we will consider the optimization problem for a two-connected elastic body in which we specify surface forces on both internal and external boundaries. In this case, (2.13.9) is not satis ed. Let us formulate this problem. Let M be a space of controls, and let a twodimensional domain q occupied by an elastic body be de ned for every q 2 M . The boundary Sq of q consists of two connected components, S2q and S1 are the internal and external boundaries of q , respectively (see Fig. 5.10.1). We prescribe \self-balanced" forces F on S1 and zero forces on S2q , S1 and F being independent of a control q, and S2q being chosen from the conditions of optimization. Let us specify the space M as follows:
e M = q j q 2 C l]+3 ( 0 2 ]) r1 < q( ) < r2

202 ]

(5.10.1)

where r1 and r2 are positive constants. For each q 2 M , let the two-connected domain q be such that the internal boundary S2q is de ned in polar coordinates by e q, and the external boundary S1 by a xed function 2 C l]+3 ( 0 2 ]). The domain

394

Chapter 5. Optimization of deformable solids

y2
F = ( F1 , F2 )

q(a )

S2 q

y1 S1

Figure 5.10.1: Domain q occupied with an elastic body \self-balanced" forces F act on S1 is de ned by (5.9.2) and the mapping Pq by (5.9.4), where Gq : E ;1 ( q ) ! E ( ) has the form
;1

(r ) ! Gq (r ) = ( ') 2 )+ ( '= : = r ; ( q( ; q( ) ) ) The inverse mapping G;1 : E ;1 ( ) ! E ;1 ( q ) has the form q ( ') ! G;1 ( ') = (r ) q r = 2q(') ; (') + (') ; q(')]
( ) 11 (u) 1q + 12 (u) 2q ) S2q 21 (u) 1q + 22 (u) 2q ) S2q ) 11 (u) 1 + 12 (u) 2 ) S1 : 21 (u) 1 + 22 (u) 2 ) S1

(5.10.2)

= ':

(5.10.3)

The operator Aq is de ned by (5.9.7) and the boundary operators Bq and B are given by the expressions ( Bq u = ( ( ( Bu = ( (5.10.4) (5.10.5)

Here, iq and i are the components of the unit outward normal to S2q and S1 , respectively, i = 1 2. De ne the spaces Vlq and Hlq in the form

Vlq = C l+2 ( q )2 Hlq = C l ( q )2 C l+1 (S2q )2 C l+1 (S1 )2

(5.10.6)

5.10. Optimization of the internal boundary

:::

395

where l > 0, l not an integer. Obviously, Lq = (Aq Bq B ) 2 L(Vlq Hlq ): ^ The kernel space of the operator Lq is the three-dimensional space Vlq = Q, where ^lq : Q is de ned by (5.9.11). The following functions form a basis in V 'q1 = (1 0) 'q2 = (0 1) 'q3 = (y2 ;y1): (5.10.7) For given functions of volume and surface forces (f R F ) de ned on q , S2q , and S1 , respectively, consider the problem: Find a function of displacements u such that Aq u = f in q Bq u = R on S2q (5.10.8) Bu = F on S1 We assume that the volume and surface forces (f R F ) are self-balanced, ^ i.e., these are orthogonal to the space Vlq in the sense that
Z
qZ

fi dy +
q

(f1 y2 ; f2 y1 ) dy + +
Z

S2q

Ri ds +

SZ 1

Fi ds = 0

i=1 2

The conditions (5.10.9) are necessary and su cient for the existence of a solution of the problem (5.10.8), see Hlavacek and Necas (1970). From (5.10.9), it follows ^ that the space Hlq = Hlq n Lq (Vlq ), where Lq = (Aq Bq B ), is three-dimensional, ^ and the following functions form a basis in Hlq : q1 = ((1 0) (1 0) (1 0)) (5.10.10) q2 = ((0 1) (0 1) (0 1)) q3 = ((y2 ;y1 ) (y2 ;y1 ) (y2 ;y1 )): Here, in the expressions for the functions qi , the rst pair belongs to C l ( q )2 , the second one to C l+1 (S2q )2 , and the third one to C l+1 (S1 )2 . In this case, the equality (2.13.23) holds with kq = 3, and we de ne the operator Tq by (2.13.24). Owing to Remark 2.13.1, we obtain the following Theorem 5.10.1 Let spaces Vlq and Hlq be de ned by (5.10.6), let an operator Lq = (Aq Bq B ) be de ned by the formulas (5.9.7), (5.10.4), (5.10.5), and let an operator Tq be given by (2.13.24) with kq = 3, where 'qi , qi are determined by (5.10.7) and (5.10.10). Then, the operator Lq1 = Lq + Tq is an isomorphism of Vlq onto Hlq :

S1

(F1 y2 ; F2 y1) ds = 0:

S2q

(R1 y2 ; R2 y1 ) ds (5.10.9)

396

Chapter 5. Optimization of deformable solids

Suppose now that the surface forces F given on the external boundary satisfy the conditions

2 C l+1 (S1 )2
Z

S1

(F1 y2 ; F2 y1 ) ds = 0:

S1

Fi ds = 0

i=1 2

(5.10.11)

Then, due to Theorem 5.10.1, there exists a unique function uq 2 C l+2 ( q )2 such that Aq uq = 0 in q Bq uq = 0 on S2q (5.10.12) Buq = F on S1 Tq uq = 0 in Hlq : In the same way as above, by using Theorem 5.10.1, we prove the subsequent assertion.

Theorem 5.10.2 Let the conditions of Theorem 5.10.1 hold, and let sets M , M1

be de ned by (5.10.1) and (5.9.17), the functionals G1 and G0 given by (5.9.15) and (5.9.16), where uq is the solution of the problem (5.10.12). Let also a function F satisfy the conditions (5.10.11), and let a nonempty set Uad be given by (5.9.18). Then, there exists a solution of the problem (5.9.19).

Obviously, in the considered case, the optimization problem consists in nding the shape of the internal boundary of the elastic body of minimal weight (volume) that satis es the constraint on strength.

5.11 Optimization problems on manifolds and shape optimization of elastic solids

We have studied earlier shape optimization problems by using the following approach. Let M be a set of controls. To each q 2 M a domain q Rn is assigned, and one considers the problem of nding a function uq de ned on q that satis es Aq uq = fq . Here, Aq is some elliptic operator acting from a space Vq to a space Hq , Vq and Hq consisting of functions de ned on q and on its boundary Sq . The optimization problem is to minimize or maximize a goal functional under some restrictions. However, in the general case, the goal and restriction functionals cannot be de ned on various Vq , q 2 M . Besides, for the existence of a solution of the optimization problem, it is necessary that the goal and restriction functionals possess some continuity with respect to a control q, but it is inconvenient for one to establish the continuity when one works with various Vq 's. This is why

5.11. Optimization problems on manifolds

:::

397

one applies a di eomorphism Pq of the set q onto a xed set , and after the change of variables corresponding to this di eomorphism one arrives to the problem A(q)u(q) = f (q) in the xed domain and on its boundary S , for all q 2 M . Now, u(q) 2 V , where V is a space of functions on , and the goal and restriction functionals are given on V . However, in some cases, the construction of the di eomorphisms Pq for all q 2 M , the passage to the problem A(q)u(q) = f (q) in the xed domain , and the solution of the problem obtained may be di cult. In this section, we will introduce and study another approach to the shape optimization, which is based on the transition to equations on the boundary, and so on the solution of an optimization problem on manifolds. In this case, instead of the di eomorphisms Pq , we should de ne maps Iq of the boundaries Sq , and the domains of the maps Iq should be the same for all q 2 M . Denoting it by T , we obtain state equations on the xed set T Rn;1 , while q Rn . Of course, such an approach may be utilized if the fundamental solution of the state equations is found. Below, we apply this approach to the optimization of elastic solids.

We consider a shape optimization problem for a two-dimensional elastic solid. Let M be a set of controls, and let to each q 2 M a bounded domain q R2 with a smooth boundary Sq be assigned. We suppose q is multiply-connected, and denote by Sq0 the external boundary of q , and by Sqi , 1 i p, the other components of Sq , where Sqi \ Sqj = ? for i 6= j , Sq0 envelops the others Sqk , and Sqk , k = 1 : : : p, do not envelop each other. Sqi is de ned by a periodic function qi : (; ) ! R2 , i = 0 1 : : : p , see Fig 5.11.1. So, we de ne a set of controls M by e M = q = (q0 q1 : : : qp ) qi = (qi1 qi2 ) 2 C m+1 ( ; ])2

Formulation of the problem

5.11.1 Optimization problem for an elastic solid

In the sequel, we consider all periodic functions as being given either on the unit circle or on R and points t + 2 k, k 2 Z, are identi ed. Thus, we consider periodic functions on R=2 Z|the factor-group consisting of classes t_ = t + 2 Z containing a point t. The set M is provided with the topology generated by that of e C m+1 ( ; ])2(p+1) :

dqi1 (t) 2 + dqi2 (t) 2 > c > 0 (5.11.1) 0 dt dt qi (t) 2 Qi t 2 (; ] i = 0 1 : : : p : Here, Qi are some open sets in R2 such that, for all q 2 M , the above condie tions on q are satis ed, C m+1 ( ; ]) is the subspace of periodic functions of C m+1 ( ; ]). m2N m 3

398

Chapter 5. Optimization of deformable solids

x2

Sq1

Sq2

Sq0

x1 q2 q0
p t

q1
-p

Figure 5.11.1: Multiply-connected domain q Sqi 's are de ned by periodic functions qi Notice that the mapping qi is a homeomorphism of (; ] onto Sqi for an arbitrary q 2 M , i = 0 1 : : : p. The operator Aq of the theory of elasticity is de ned by

Aq = ; u ; ( + ) grad div u

in q :

(5.11.2)

Here, u = (u1 u2) is a vector function of displacements, , are positive constants. We denote by "(u) = ("ij (u)), (u) = ( ij (u)) the strain and stress tensors,

@u "ij (u) = 1 @x i + @uj 2 j @xi ij (u) = div u ij + 2 "ij (u)

i j=1 2

(5.11.3)

where ij is the Kroneker delta. The traction operator (the operator of surface forces) Tq is de ned on Sq by

Tq u = ((Tq u)1 (Tq u)2 ) (Tq u)i = ij (u) jq on Sq i j = 1 2:

(5.11.4)

Here and below, the summation over repeated index is implied, jq are the components of the unit outward normal q to Sq . Various formulations of problems of the theory of elasticity may be considered, in particular, displacement, traction, mixed, and other ones (see, e.g., Kupradze et al. (1979)). We will be engaged in the traction formulation

Tq u = Fq

on Sq :

(5.11.5)

5.11. Optimization problems on manifolds

:::

399

So we consider the problem: Find a function uq satisfying

Aq uq = 0 Tq uq = Fq

in q on Sq :

(5.11.6) (5.11.7)

The case when the function of body forces, i.e., the right hand side of (5.11.6), is not equal to zero, may be reduced to the problem (5.11.6), (5.11.7). Further, we suppose that the boundary Sq0 is xed, i.e.,

Sq0 = S

q2M

(5.11.8)

the surface forces Fq are not equal to zero only on S , and they are xed and self-balanced:

Fq Sqi = 0 i = 1 ::: p 2 Fq S = (F1 F2 ) 2 H m; 3 (S )2 Z Fi ds = 0 i=1 2

ZS

(5.11.9)

(F1 x2 ; F2 x1 ) ds = 0:

We introduce spaces

Vmq = H m ( q )2 3 Hmq = H m;2 ( q )2 H m; 2 (Sq )2 :

(5.11.10)

Then, the operator Gq = (Aq Tq ) is a linear continuous mapping of Vmq into Hmq , i.e., Gq 2 L(Vmq Hmq ), and by known results, see Theorem 2.13.1 and Section 5.10, we obtain

Theorem 5.11.1 Let a set M be de ned by (5.11.1), and let (5.11.8), (5.11.9) hold. Then, for each q 2 M , the following representations are valid
^ Vmq = Vmq Vmq ^ Hmq = Hmq Hmq ^ ^ where Vmq = ker Gq , Hmq = Gq (Vmq ), Gq = (Aq Tq ). The subspaces Vmq and ^ mq are three-dimensional, and '1 = (1 0), '2 = (0 1), '3 = (x2 ;x1 ) is a basis H ^ in Vmq , 1 = ((1 0) (1 0)), 2 = ((0 1) (0 1)), 3 = ((x2 ;x1 ) (x2 ;x1 )) is a ^ basis in Hmq . For each q 2 M , there exists a unique uq 2 Vmq satisfying (5.11.6), (5.11.7).

400
Z

Chapter 5. Optimization of deformable solids

We introduce the following functionals on M :

0 (q ) = 1 (q ) = max juq (x)j ; c1
q

dx uq 2 Vmq
2 Xh

x2 q

x2 q i j =1 Z 3 (q ) = ij (uq )"ij (uq ) dx ; c3

q

2 (q ) = max

(5.11.11) i2 1; ij (uq )(x) ; 2 11 (uq )(x) + 22 (uq )(x) ij ; c2

where c1 c2 c3 are positive constants. Note that other functionals on M may also be considered. Now, suppose that

M1 is a compact subset of M:
In particular, the set M1 may be de ned by
e M1 = q = (q0 q1 : : : qp ) 2 M qi (t) 2 Qi t 2 (; ] e Qi are closed subsets in Qi kqi kC m+1+ ( ; ]) c 2 (0 1] i = 1 : : : p :

(5.11.12)

We remind that q0 is considered to be xed (see (5.11.8)), C k+ ( ; where k 2 N , denotes the Holder space with the norm

]),

kukC k+

(;

]) = kukC k ( ;

]) +

tt2;

sup 0
i (q )

dk u (t) ; dk u (t0 ) : dtk ] jt ; t0 j dtk

1 0 i=1 2 3
0 (q ):

De ne a set of admissible controls Mad as follows

Mad = q j q 2 M1 q 2 Mad ~

(5.11.13) (5.11.14)

and consider the optimization problem: Find q satisfying ~

q inf 0 (~) = q2M

ad

From the physical point of view, the problem (5.11.14) corresponds to the minimization of the area (weight) of the elastic solid under restrictions on displacements, stresses, and strain energy. Other restrictions of the form k (q) 0 may also be considered.

5.11. Optimization problems on manifolds

:::

401

State equations on the boundaries and on the unit circle

Au = ; u ; ( + ) grad div u = 0 G(x y) is the symmetric tensor de ned by Gij (x y) = c1 c2 ij log r ; (xi ; yi )(2xj ; yj ) r
where

Let G(x y) = (Gij (x y)) be the tensor of fundamental solutions of the equation

u = (u1 u2 ): i j=1 2

(5.11.15) (5.11.16)

i=1 (5.11.17) 1 c1 = ; 8 (1 ; ) c2 = 3 ; 4 = 2( + ) : From the physical point of view, the function Gij (x y) de nes a displacement ui (x) engendered by the unit force Pj (y) concentrated at a point y and directed along the xj axis. By Bk (x y) = (Bijk (x y)) and Tk (x y) = (Tijk (x y)) we denote the strain and stress tensors at a point x that are engendered by the force Pk (y). Due to (5.11.3) and (5.11.16), we obtain h i B (x y) = c1 (1 ; 2 )( + ); +2 (5.11.18) ik j jk i ij k r2 i j k r2 h i c Tijk (x y) = r3 c4 ( ik j + jk i ; ij k ) + r22 i j k 2 1 i = xi ; yi c3 = ; 4 (1 ; ) c4 = 1 ; 2 : ijk

r=

2 X

(xi ; yi )2

1 2

(5.11.19)

The force t(x) = (t1 (x) t2 (x)) at a point x of a surface with a unit outward normal = ( 1 2 ) is de ned by ti (x) = ij (x) j (x). So, denoting by Rik (x y) the value at a point x of the i-th component of the surface force generated by Pk (y), we get due to (5.11.19) : (5.11.20) R ( x y ) = c3 c ( ; ) + c + 2 i k
ik

r2

4 k i

i k

4 ik

r2

j j

Let u = (u1 u2) be a smooth function satisfying (5.11.15) in q . By Betti's formula, see Banerjee and Butter eld (1981), we obtain

ui (x) =

Sq

(Tq u)j (y)Gij (x y) ; Fqij (x y)uj (y) dSy

x2 q i=1 2

(5.11.21)

402

Chapter 5. Optimization of deformable solids

where Fqij (x y) = Rji (y x) for = q , i.e.,

Butter eld (1981), the representation formula (5.11.21) yields the following expression 1 u (x) = Z (T u) (y)G (x y) ; F (x y)u (y) dS q j ij qij j y 2 i Sq (5.11.23) x 2 Sq i = 1 2: The integral from the second addend in (5.11.23) should be understood in the sense of the Cauchy principal value. Considering the problem (5.11.6), (5.11.7), we obtain the equation

c h Fqij (x y) = r3 c4 ( qi (y)(yj ; xj ) ; qj (y)(yi ; xi )) 2 i + c4 ij + 2(yi ; xir)(yj ; xj ) (yk ; xk ) qk (y) : (5.11.22) 2 Because of the jump relation for Fq (x y) = (Fqij (x y)) on Sq , see Banerjee and

Nq u(x) = u(x) + 2
where

Sq

Fq (x y)u(y) dSy = fq (x)

Z

x 2 Sq

(5.11.24) (5.11.25)

fq (x) = 2

The boundary Sq is de ned by the function q = (q0 q1 : : : qp ) 2 M (see Fig. 5.11.1). Denoting vqi (t) = u(qi (t)), i = 0 1 : : : p, vq = (vq0 : : : vqp ), we transform the problem (5.11.24) into the following one (P (q)vq )i (t) = vqi (t) + 2 = gi (t) Here,
2 2 2 D(qj )( ) = dqj1 ( ) + dqj2 ( ) (5.11.27) d d gi (t) = fq (qi (t)) and gi is independent of q because Fq is not equal to zero only on Sq0 , and Sq0 is xed, see (5.11.8), (5.11.9). In the case when i = j , the integral in (5.11.26) should
1

Sq

G(x y)Fq (y) dSy :

p XZ j =0

t2 ;

Fq (qi (t) qj ( ))vqj ( )D(qj )( ) d

] i = 0 1 : : : p:

(5.11.26)

be understood as the limit of the integrals

Z
;

Fq" (t )vqi ( )D(qi )( ) d

5.11. Optimization problems on manifolds

(

:::

403

Fq" (t ) = Fq (qi (t) qi ( )) if jt ; j " 0 if jt ; j < "

as " tends to zero. De ne a space Wm as follows

5.11.2 Spaces and operators on R =2 Z, auxiliary statements

e 1 Wm = u = (u0 u1 : : : up ) ui = (ui1 ui2 ) 2 H m; 2 (; i = 0 1 ::: p m 3 : e By H s (; ), s > 0, we denote the subspace of H s (; functions. The norm in Wm is de ned by

)2

(5.11.28)

) consisting of periodic
1 !2

kvkm =
Here,

p 2 XX i=0 k=1
"

kuik k2e m; 1 (; H 2

(5.11.29)

kvkH m; 2 (; e 1

)=

where an , bn are the Fourier coe cients of the function v, i.e.,

1 a0 + X(a cos nt + b sin nt) v(t) = 2 n n

X a2 + (a2 + b2 )n2m;1 0 n n n=1

1

#1 2

(5.11.30)

an = 1

Z n=1
; ;

Z bn = 1 v(t) sin nt dt:

v(t) cos nt dt

(5.11.31)

Let ^ ^ Wmq = ( q Vmq ) q : (5.11.32) ^ Here, Vmq is ker(Aq Tq ) (see Theorem 5.11.1), q is the trace operator on Sq , q is ^ the function de ned in (5.11.1). It follows from Theorem 5.11.1 that Wmq is the 3 , three-dimensional subspace of Wm with the basis f'qi gi=1

'qi = ('qi0 'qi1 : : : 'qip ) i=1 2 3 'q1k = (1 0) 'q2k = (0 1) 'q3k = (qk2 ;qk1 ) k = 0 1 ::: p

(5.11.33)

404

Chapter 5. Optimization of deformable solids

qki de ned in (5.11.1). We denote by Jq the operator of the simple layer potential
(Jq v)(x) = and let
qi = (Jq i )
1 = (1 0) 2 = (0 1) 3 = (x2 ;x1 )

Sq

G(x y)v(y) dSy

x 2 Sq

(5.11.34)

i=1 2 3 on Sqi on Sqi on Sqi i = 0 1 : : : p: i=1 2 3

(5.11.35)

By (5.11.34), (5.11.35), we get

qij (t) = qi = ( qi0 qi1 p XZ k=0
;

: : : qip )

G(qj (t) qk ( )) i (qk ( ))D(qk )( ) d

j = 0 1 ::: p
(5.11.36)

D(qk ) de ned by (5.11.27).

Theorem 5.11.2 Let a set M be de ned by (5.11.1), and let (5.11.8), (5.11.9) hold. Let also an operator P (q) be de ned by (5.11.26). Then, P (q) 2 L(Wm ), and for each q 2 M the following representations are valid ^ ^ Wm = Wmq Wmq = Emq Emq (5.11.37)
^ ^ where Wmq = ker P (q), Emq = P (q)(Wm ), Emq is the three-dimensional subspace 3 de ned by (5.11.36). There exists a unique vq 2 Wmq of Wm with the basis f qi gi=1 satisfying (5.11.26).

Jq is an isomorphism of H m; 2 (Sq )2 onto H m; 2 (Sq )2 , see Chudinovich (1991), Wendland (1985). It is obvious that, if uq is a solution of the problem (5.11.6), (5.11.7), then u = q uq is a solution of the problem (5.11.24). On the contrary, if u is a solution of the problem (5.11.24), then the function uq (x) =
Z

Proof. The operator Jq is a pseudodi erential operator of order ;1 on Sq , and 3 1

Sq

G(x y)Fq (y) ; Fq (x y)u(y) dSy

x2 q

is a solution of the problem (5.11.6), (5.11.7). The problem (5.11.26) is obtained from (5.11.24) by the change of variables corresponding to the one-to-one mapping e q 2 C m+1 (; )2(p+1) . So, Theorem 5.11.2 follows from Theorem 5.11.1.

5.11. Optimization problems on manifolds

:::

405

Remark 5.11.1 The space Wmq is de ned nonuniquely by (5.11.37), and so we

^ de ne Wmq so that it is orthogonal to Wmq with respect to the scalar product in 2(p+1) . L2 (; ) ^ De ne an operator ;q1 2 L(Wm Wmq ) by ^ ;q1 is the orthogonal projection of Wm onto Wmq with re2(p+1) . spect to the scalar product in L2(; ) ^ ^ Also, de ne an operator ;q2 2 L(Wmq Emq ) as follows
3 X 3 X

(5.11.38)

u=

Theorem 5.11.3 Let a set M and an operator Pq be de ned by (5.11.1) and (5.11.26), respectively. Then, the operator Nq = P (q)+;(q), where ;(q) = ;q2 ;q1 ,
is an isomorphism of Wm onto itself. ^ ^ Proof. Obviously, ;(q) 2 L(Wm Emq ) and ;(q)(Wm ) = Emq . Therefore, by Theorem 5.11.2, the operator P (q)+;(q) is a continuous bijection from Wm onto Wm , and so by the Banach theorem P (q) + ;(q) is an isomorphism of Wm onto Wm .

i=1

ci 'qi

;q2 u =

i=1

ci qi :

(5.11.39)

Lemma 5.11.1 Let a set M and an operator ;(q) = ;q2 ;q1 be de ned by (5.11.1), (5.11.38), (5.11.39), and let qn ! q0 in M (M equipped with the topology e generated by the C m+1 (; )2(p+1) topology ). Then, k;(qn ) ; ;(q0 )kL(Wm Wm ) ! 0 as n ! 1. Proof. Let u = (u0 u1 : : : up) 2 Wm . By (5.11.38), we have
k;qn 1 u ; uk2 2(; L
)2(p+1)

= min ci

p X k=0

kuk ;

3 X

i=1

ci 'qn ik k2 2 (; L

)2 :

(5.11.40)

We denote by cin a solution of the problem (5.11.40), i.e., cin minimizes the norm in (5.11.40). Then, we get

cin

3 p XX

i=1 k=0

('qn ik 'qn jk )L2 (;

)2

p X k=0

(uk 'qn jk )L2 (;

)2

j = 1 2 3:
(5.11.41)

The matrix of the system (5.11.41) is nondegenerate because the functions ^ f'qn i g3=1 form a basis in Wmqn . Therefore, cin are de ned uniquely. Let now i n ! q 0 in M . Then 'qn ik ! 'q0 ik in C m+1 (; )2 , and so cin ! ci0 as n ! 1, e q where ci0 is a solution of the problem (5.11.41) for n = 0. As the embedding of

406

Chapter 5. Optimization of deformable solids

e C m+1 (; ) into L2 (; ) is compact, we obtain that cin ! ci0 uniformly in u 2 K , where K is an arbitrary bounded set in M . e It may be shown that qn ij ! q0 ij in H m(; )2 as qn ! q0 in M . n ) ! ;(q 0 ) in L(Wm ), concluding the proof of the lemma. Therefore, ;(q In the space Wm , we de ne an operator N (q) as follows

v = (v0 v1 : : : vp ) 2 Wm
(N (q)v)i (t) = By (5.11.26), we have
p XZ j =0

N (q)v = f(N (q)v)i gp=0 i

(5.11.42)

t2 ;

Fq (qi (t) qj ( ))vj ( )D(qj )( ) d

]

i = 0 1 : : : p:
(5.11.43)

P (q) = I + 2N (q)
where I is the identity operator in Wm .

Lemma 5.11.2 Let a set M be de ned by (5.11.1) and let an operator N (q) be de ned by (5.11.42), (5.11.22), (5.11.27). Then, N (q) 2 L(Wm ) and the function N : q ! N (q) is a continuous mapping of M into L(Wm ). Proof. By (5.11.22), we have
Fq (qi (t) qj ( )) = (Fqkn (qi (t) qj ( )))2 n=1 k c3 nc Fqkn (qi (t) qj ( )) = ( t) 4 qk (qj ( ))(qjn ( ) ; qin (t)) ; qn (qj ( ))(qjk ( ) ; qik (t)) q + c4 kn + 2(qjk ( ) ; qik (t())(t)jn ( ) ; qin (t))
2 X

s=1

(qjs ( ) ; qis (t)) qs (qj ( )) :

(5.11.44)

Here, ( t) =
q (qj ( )) =
2 X

a( ) dqj2 ( ) ;a( ) dqj1 ( ) d d 1 2 dqj1 ( ) + dqj2 ( ) 2 ; 2 : a( ) = d d

s=1

(qis (t) ; qjs ( ))2 (5.11.45)

5.11. Optimization problems on manifolds

:::

407

In the case when i 6= j , the elements Fqkn (qi (t) qj ( )) are nonsingular and they generate smoothing operators. So, we consider the case when i = j . As e qi 2 C m+1 ( ; ])2 , by (5.11.45) and the Taylor formula, we get
2 X

Therefore, the terms containing A( t) as a factor in (5.11.44) generate a nonsingular operator. The two remained addends in (5.11.44) are singular and similar. So, it is su cient to show that 9 the operator R(q) de ned by > > > > 1 m; 2 (; ), > e > u2H > > > > > (R(q)u)(t) = Z : 1 > = > > ( t) qk (qi ( ))(qin ( ) ; qin (t))D(qi )( )u( ) d > ; > > > 1 > e m; 2 (; ) into itself, and if q ! q 0 > is a continuous mapping of H > > 0 ) in L(H m; 1 (; )): e 2 in M , then R(q ) ! R(q (5.11.46) By the Taylor formula we obtain 0 qin ( ) ; qin (t) = qin (t)( ; t) + aq ( t) (5.11.47) ( t) =
2 X

s=1

(qis ( ) ; qis (t)) qs (qi ( )) = jA( t)j c(t ; )2 :

s=1

(qis ( ) ; qis (t))2 =

2 X

Substituting (5.11.47), (5.11.48) into R(q)u, we get R(q)u = R1 (q)u + R2 (q)u (R1 (q)u)(t) = (R2 (q)u)(t) =
2 X

jaq ( t)j cj ; tj2

s=1

0 qis (t)2 ( ; t)2 + q ( t)

(5.11.48) (5.11.49) (5.11.50) (5.11.51) (5.11.52)

j q ( t)j c1 j ; tj3 :
Z

Z s=1
;

0 qis (t)2

;1

0 qin (t)

wq ( )u( ) d ;t ;

H ( t)wq ( )u( ) d :

Here, ; is the unit circle on the x1 Ox2 plane centered at 0, the origin of the curvilinear coordinate on ; is a point of ; lying on Ox1 , the direction of count is counter-clockwise, wq ( ) = qk (qi ( ))D(qi )( ) e H ( t) = fq ( t) q ( t)

408

Chapter 5. Optimization of deformable solids

eq ( t) = fq ( t) =

2 X

s=1 2 X s=1

0 0 qis (t)2 aq ( t)( ; t) ; qin (t)bq ( t) 0 qis (t)2 ( ; t)3 +

2 X

Obviously, R3 u is the convolution on ; of the principal value of the distribution 1 and ;wq u. We denote by ck (f ) the Fourier coe cients of a function f for the t complex form of the Fourier series, Z f (t)e;ikt dt: ck (f ) = 21 ; Then, we have Schwartz (1961)
1 X

Notice that the function ! ;t is discontinuous on ; at the point 0 = t+ . De ne an operator R3 as follows Z ) (R3 u)(t) = wq ( ;ut( ) d : (5.11.54)
;

s=1

0 qis (t)2 q ( t)( ; t):

(5.11.53)

k=;1

jck (R3 u)j2 = 4

1 2 X jc ( 1 )c (;w u)j2 k t k q k=;1

(5.11.55) ]), the (5.11.56)

e Since jck ( 1 )j const for all k (see Edwards (1979)) and wq 2 C m ( ; t Parseval equality and (5.11.55) yield

kR3 uk2 2(; L

For a 6= 0, we get

)=2

1 X

k=;1

jck (R3 u)j2 ckuk2 2(; L

):

a;1 (R3 u)(t + a) ; (R3 u)(t) Z wq ( )u( ) ; wq ( )u( ) d ;1 =a ;t;a ;t Z; ) = a;1 wq ( + a)u( + at ; wq ( )u( ) d : ;
;

(5.11.57) ), then

e By (5.11.56), (5.11.57), we obtain that, if u 2 H 1 (; Z

d ;1 d dt (R3 u) (t) = ; ( ; t) d (wq ( )u( )) d

and

kR3 ukH 1 (;

c1 kukH 1(;

):

5.11. Optimization problems on manifolds

:::

409

By analogy, we have

kR3 ukH m (;

c1 kukH m(;

and taking into account (5.11.51), (5.11.54) we get )). (5.11.58) By using the interpolation (see Lions and Magenes (1972)), we obtain that (5.11.58) holds for m replaced with m ; 1 . The operator R2 is smoothing and 2 also satis es (5.11.58). Now, taking into account that q is a periodic and smooth function on R, we obtain (5.11.46), concluding the proof. 2 Sp We remind that by q we denote a domain in R with a boundary Sq = i=0 Sqi , Sq de ned by a function q = (q0 q1 : : : qp ) 2 M . Theorem 5.11.4 Let a set M be de ned by (5.11.1) and let qn ! q0 in M . Then, for each su ciently large n, there exists a mapping Pn such that Pn is a C m -di eomorphism of q0 onto qn and Pn ! I in C m ( q0 )2 , where I is the identity operator in q0 : Proof. First, we will prove the theorem for the case when qn and q0 are simply connected. Let Sn and S0 be the boundaries of qn and q0 , respectively, let be a small positive number, and let G0 = x j x 2 q0 yinf kx ; yk > : (5.11.59) 2S
0

e q ! R1 (q) is a continuous mapping of M into L(H m (;

Let also T0 be the boundary of G0 and F0 = q0 n G0 , see Fig. 5.11.2. By s we denote points of a parameterization of T0. Then, we can consider that s 2 R=2 Z. Outside of G0 , we de ne curvilinear coordinates (s r), the r axis at a point s is normal to T0 . Let now e qn ! q0 in C m+1 ( ; ])2 : (5.11.60) Then, for su ciently large n, we have Sn \ T0 = ?, and we de ne a function fn : R=2 Z 0 ] ! R2 as follows

fn (s r) = (s n (s r)) @ n (s 0) = 1 @r

n (s r) =

m+1 X k=1

ank (s)rk

(5.11.61)

@ k n (s 0) = 0 k = 2 ::: m (5.11.62) @rk Q(s n (s )) 2 Sn : Here, Q is the transformation of curvilinear coordinates (s r) into Cartesian coordinates (x1 x2 ). By (5.11.61), (5.11.62), we get an1 (s) = 1 ank (s) = 0 k = 2 ::: m (5.11.63)

410

Chapter 5. Optimization of deformable solids

and (5.11.60) yields an(m+1) (s) ! 0 uniformly in s. Therefore, fn is a one-to-one mapping. Since Sn 2 C m+1 , we obtain from (5.11.62) that
e an(m+1) 2 C m ( ;
r s

]):

(5.11.64)

x2 S0 T0

G0

F0

0
Sn

x1

Figure 5.11.2: Domains qn , q0 and G0 Now, we de ne a mapping Pn : q0 ! R2 as follows: if x 2 G0 Pn (x) = x ;1 )(x) (Q fn Q if x 2 q0 n G0 :

(

(5.11.65)

By (5.11.61){(5.11.65), we get that Pn is a C m -di eomorphism of q0 onto qn and Pn ! I in C m ( q0 )2 . In the case when qn and q0 are multiply-connected, we introduce curvilinear coordinates (s r) in a vicinity of each component of the boundary of the domain q0 , the functions fn are de ned in each vicinity, and by analogy with (5.11.65), we de ne a C m -di eomorphism of q0 onto qn . Thus, the theorem is proven.

5.11.3 Optimization problem on R =2

State equations and functionals By Theorem 5.11.2, for each q 2 M , there exists a unique vq satisfying vq 2 Wmq P (q)vq = g (5.11.66)
where the operator Pq and g = (g0 g1 : : : gp ) are de ned by (5.11.25){(5.11.27). So, the function M 3 q ! vq 2 Wm is well de ned.

5.11. Optimization problems on manifolds

:::

411

Theorem 5.11.5 Let a set M be de ned by (5.11.1), and let (5.11.8), (5.11.9) hold. Then, the function q ! vq de ned by the solution of the problem (5.11.66) is a continuous mapping of M into Wm . Proof. We de ne a mapping J : M Wm ! Wm by J (q u) = (P (q) + ;(q))u ; g
(5.11.67)

and by we denote the implicit function de ned by (q) 2 Wm , J (q (q)) = 0. Since ;(q)vq = 0 (see (5.11.38) and Remark 5.11.1), we obtain J (q vq ) = 0, i.e., (q) = vq . By Lemmas 5.11.1, 5.11.2 and by (5.11.43), the function q ! P (q)+;(q) is a continuous mapping of M into L(Wm ). By Theorem 5.11.3, the operator P (q) + ;(q) is an isomorphism of Wm onto itself. Now, the theorem follows from the implicit function theorem (Theorem 1.9.1). Let

qn ! q0

in M:

(5.11.68)

We denote by un and u0 the solutions of the problem (5.11.6), (5.11.7) for q = qn and q = q0 , respectively. By (5.11.21) we have
p Z n (x) = ; X u j =0 ;

Fq (x qjn ( ))vqn j ( )D(qjn )( ) d + P (x) Fq (x qj0 ( ))vq0 j ( )D(qj0 )( ) d + P (x)

x 2 qn
(5.11.69)

u0 (x) = ;

p XZ j =0
;

x 2 q0
(5.11.70) (5.11.71)

P (x) =

Z
;

G(x q0 ( ))F (q0 ( ))D(q0 )( ) d :

Here, q0 is the rst component of an element q 2 M , q0 is xed, F = (F1 F2 ), see (5.11.8), (5.11.9), vqn j , vq0 j are the solutions of the problem (5.11.26) for q = qn and q = q0 , respectively. By Theorem 5.11.4, for each large n, there exists a C m -di eomorphism Pn of q0 onto qn . So, we de ne a function un as follows: ~

un (x) = un (Pn (x)) ~

(5.11.69){(5.11.72).

u n 2 H m ( q 0 )2 : ~

(5.11.72)

Theorem 5.11.6 Let a set M be de ned by (5.11.1), and let (5.11.8), (5.11.9), and (5.11.68) hold. Then, un ! u0 in H m( q0 )2 , where un and u0 are de ned by ~ ~

412

Chapter 5. Optimization of deformable solids

functions on q0 by

e 1 Proof. De ne operators Un mapping the space H m; 2 (;

)p+1 into a space of

n qjs ( ) ; Pn (x)s uj ( ) d (5.11.73) nj (x ) j =0 ; e 2 u = (u0 u1 : : : up ) 2 H m; 1 (; )p+1 x 2 q0 n = 0 k k + 1 k + 2 : : :

(Un u)(x) =

p XZ

Here, s = 1 or 2, P0 is the identity operator in q0 , Pn (x)s is the s-th component of the vector Pn (x),
nj (x ) =
2 X n (qji ( ) ; Pn (x)i )2 : i=1

(5.11.74) (5.11.75) (5.11.76)

To prove the theorem, it is su cient to show that

e 1 Un 2 L(H m; 2 (;

)p+1 H m ( q0 )) as n ! 1:

and

kUn ; U0 kL(H m; 1 (; e 2

)p+1 H m ( q0 )) ! 0

Let Sq0 j be an arbitrary component of the boundary Sq0 . In a vicinity of Sq0 j in q0 , we introduce curvilinear coordinates t, r, i.e., we introduce the mapping ; ] 0 ] 3 (t r) ! Q(t r) = (Q1 (t r) Q2 (t r)) = x the function t ! Q(t 0) maps ; ) onto Sq0 j and Q is a C m -di eomorphism of e ; ) 0 ] onto Q( ; ) 0 ]) q0 , Q( r) 2 C m ( ; ]) for any xed r 2 0 ], and is a small positive constant. De ne an operator Mn by Z q n ( ) ; P (Q(t r)) n s js (Mn u)(t r) = (Q(t r) ) u( ) d nj (5.11.77) ; e m; 1 (; ) (t r) 2 ! = (; ) (0 ): 2 u2H In order to prove (5.11.75), (5.11.76), it is su cient to show that
e 1 Mn 2 L(H m; 2 (;

kMn ; M0 kL(H m; 1 (; e 2

) H m (!)) as n ! 1: ) H m (!)) ! 0

(5.11.78) (5.11.79) (5.11.80)

By the Taylor formula, we obtain n qjs ( ) ; Pn (Q(t r))s = a(qjn )( ; t) + b(qjn )r + n ( t r) n n nj (Q(t r) ) = e(qj )( ; t)2 + f (qj )r2 + n ( t r):

5.11. Optimization problems on manifolds

:::

413

Here,

a, b, e, f are continuous functionals on e C m+1 ( ; ])2 , e(qjn ) > 0, f (qjn ) > 0 for all n, j n ( t r)j c(j ; tj2 + r2 ) j n ( t r)j c(j ; tj3 + r3 ):

(5.11.81) (5.11.82) (5.11.83) (5.11.84)

By (5.11.77), (5.11.80), and (5.11.81), we get the representation

Mn = Mn1 + Mn2
where (Mn1 u)(t r) =
Z
;

a(qjn )( ; t) + b(qjn )r e(qjn )( ; t)2 + f (qjn )r2 u( ) d

) H m (!)):

and Mn2 is a smoothing operator. It follows from Agmon et al. (1959) that
e 2 Mn1 2 L(H m; 1 (;

Let R+ stand for the set of positive numbers. Consider the function
e 2 R2 R2 3 y = (y1 y2 y3 y4 ) ! H(y) 2 L(H m; 1 (; ) H m (!)) + Z y3 ( ; t) + y4r (H(y)u)(t r) = y ( ; t)2 + y r2 u( ) d :
;

e 2 The function H is a continuous mapping of R2 R2 into L(H m; 1 (; + So, by (5.11.68), (5.11.81), we have

) H m (!)).

kMn1 ; M01 kL(H m; 1 (; e 2

) H m (!)) ! 0

as n ! 1:

Analogous results hold for the operator Mn2 , hence (5.11.75), (5.11.76) are fullled.

Theorem 5.11.7 Let a set M be de ned by (5.11.1), and let (5.11.8), (5.11.9) hold. Let also functionals i , i = 0 1 2 3, be de ned by (5.11.11), where uq 2 Vmq is the solution of the problem (5.11.6), (5.11.7). Then, i are continuous functionals on M: Proof. The continuity of the functional 0 is obvious. Let us prove the continuity of the functional 2 . The proof for the other functionals is analogous. Let qn ! q0 in M . By Theorem 5.11.6, we have un ! u0 in H m ( q0 )2 : ~ (5.11.85)

414 have

Chapter 5. Optimization of deformable solids

By the change of variables corresponding to the C m -di eomorphism Pn we

2 (q 2 2 n ) = max X ~ij (~n )(x) ; 1 ; ~11 (~n )(x) + ~22 (~n )(x) ij ; c2 u u u 2 x2 q0 i j =1 ~ij (~n )(x) = ij (un )(Pn (x)) x 2 q0 : u

Due to (5.11.85), (5.11.86), we get 2 (qn ) ! 2 (q0 ). Theorem 5.11.8 Let a set M be de ned by (5.11.1), and let (5.11.8), (5.11.9), (5.11.12) hold. Let also functionals i , i = 0 1 2 3, and let a set Mad be de ned by (5.11.11), (5.11.13), Mad being nonempty. Then, there exists a solution of the problem (5.11.14). Proof. By (5.11.12) and Theorem 5.11.7 we get that Mad is a compact set in M , The functional 0 is continuous on M , and so there exists a solution of the problem (5.11.14). For the sensitivity analysis and for construction of approximate solutions of the problem (5.11.14), it is useful to calculate the derivative of the function q ! vq , where vq is a solution of the problem (5.11.66). So, we consider this problem. We introduce the notation (q) = vq : (5.11.87) It follows from the proof of Theorem 5.11.5 that is the implicit function de ned by (q) 2 Wm , J (q (q)) = 0, J (q u) is determined by (5.11.67). By applying (5.11.44) and the representation (5.11.50){(5.11.53), it may be shown that q ! P (q) is a continuously Frechet di erentiable mapping of M into L(Wm ). It may also be shown that q ! ;q is a continuously Frechet di erentiable mapping of M into L(Wm ). Denote by P 0 (q), ;0 (q) the derivatives of these mappings at a point q. So, by applying Theorem 1.9.2, we obtain that is a continuously Frechet di erentiable mapping of M into Wm and its derivative at a point q 2 M is given by 0 (q )h = ;(P (q ) + ;(q ));1 (P 0 (q ) + ;0 (q ))h (q )
e where h = (h0 h1 : : : hp ), hi = (hi1 hi2 ), hi 2 C m+1 ( ; zero since the component Sq0 is xed (see (5.11.8)).

(5.11.86)

])2 , h0 is equal to

5.12 Optimization of the residual stresses in an elastoplastic body

We have considered optimization problems in which the state of the object is described by linear operator equations. Now, we will be engaged in a problem of

5.12. Optimization of the residual stresses in an elastoplastic body

415

optimization of the residual stresses in a solid, supposing that the state of the solid is described by nonlinear equations. Stresses and deformations are called residual if they stay inside of the solid after that the action which induced them is removed. The elds of volume and surface forces, temperature patterns, phase transformations may be an action causing the residual stresses and deformations, which appear in almost all cold and hot workings of materials. In order to decrease stresses in a construction in the process of its operation, one tries to create, in the process of the production of the construction, the residual stresses that are equal in absolute value but oppositely directed to the stresses arising in the operation. In many cases, one aims for decreasing the residual stresses and deformations. So, various problems of modeling and optimization of them appear, see, e.g., Pozdeev et al. (1982), Grigoliuk et al. (1979), Shablii and Medynskii (1981). Below, we consider a problem of optimization of the residual stresses and deformations in a tree-dimensional elastoplastic body.

5.12.1 Force and thermal loading of a nonlinear elastoplastic body

Basic equations
Supposing that deformations are small, we accept, for the process of loading, the following relation between stresses, strains, and temperature (see Shevchenko (1970))
ij (u ) = ( )("(u) ; ( ) ) ij + 2g (I (u) )eij (u)

i j = 1 2 3:

(5.12.1)

Here, ij (u ) are the components of the stress tensor, which depend on a vector function of displacements (deformations) u = (u1 u2 u3 ) and temperature , is the coe cient of linear expansion, the compression modulus, , depending on the temperature , 1 "(u) = 3
3 X @ui

1 @xi = 3 div u i=1

(5.12.2)

eij (u) are the components of the deviator of the strain tensor, eij (u) = "ij (u) ; "(u) ij "ij are the components of the strain tensor @u "ij (u) = 1 @x i + @uj 2 j @xi
(5.12.4) (5.12.3)

416

Chapter 5. Optimization of deformable solids

I (u) is the second invariant of the deviator of the strain tensor I (u) =
3 X

i j =1

(eij (u))2

(5.12.5)

We assume that, before power and thermal loading, (x) = 0 for all x 2 , where is the domain occupied by the body. We could consider that, before loading, (x) = 0 = const. Then, in (5.12.1), is the di erence between the temperature and 0 . We suppose that is a bounded domain in R3 with a Lipschitz continuous boundary S , and S = S 1 S 2 , S1 \ S2 = ?, S1 and S2 open nonempty subsets of S. If the temperature pattern in the medium is known, we are given volume forces K = (K1 K2 K3 ) and surface forces F = (F1 F2 F3 ) which act in and on S2 , respectively, and the body is fastened at S1 , then the problem of nding the function of displacements of points of the body u = (u1 u2 u3 ) reduces to the following problem:
3 X @ ij (u(x) (x)) + Ki (x) = 0 @xj j =1 3 X u S1 = 0 ij (u ) j S2 = Fi j =1

g is the plasticity modulus, which depends on I (u) and .

in

i = 1 2 3:

(5.12.6)

Here, = (

1 2 3)

is the unit -outward normal vector to S2 .

De ne a space H as follows: H = u j u = (u1 u2 u3) 2 (W21 ( ))3 u S1 = 0 : Lemma 5.12.1 In H , the norm k kH de ned by the expression

Generalized solution

(5.12.7) (5.12.8)

kukH =

("(u))2 + I (u) dx

1 2

is equivalent to the norm of the space (W21 ( ))3 : Proof. By using (5.12.3), we have
Z X 3

i j =1

("ij (u))2 dx =

Z X 3

i j =1

(eij (u))2 + 2eij (u)"(u) ij + ("(u))2 ij dx: (5.12.9)

5.12. Optimization of the residual stresses in an elastoplastic body

417

By (5.12.2){(5.12.4), we have
3 X

i j =1

eij (u) ij =

3 X

i=1

@ui ; 1 div u = 0: ii @xi 3

Z ;

(5.12.10)

From (5.12.5), (5.12.9), and (5.12.10), it follows that

Z X 3

i j =1
Z

("ij (u))2 dx =

I (u) + 3("(u))2 dx: u2H

(5.12.11)

Taking into account the Korn inequality, we obtain from (5.12.11) that ("(u))2 + I (u) dx ckuk2W21( (
))3

(5.12.12)

where c is a positive constant. The inequality opposite to (5.12.12) is obvious, so the lemma is proven. Suppose that the function g (see (5.12.1)) satis es the following conditions: I: The function g( ) is de ned and continuous on the half-plane Q = ( ) j 2 R+ 2 R R+ = y j y 2 R y 0 : II: There exist positive constants a1 , a2 , a3 such that the following estimates hold: a1 g( ) a2 ( )2Q (5.12.13) 2 ) ; g ( 2 ) ( ; ) a ( ; )2 g( 1 1 (5.12.14) 2 1 2 3 1 2 2
1 2

2 R+

2 R:

From the physical point of view, the estimate (5.12.13) means that the plasticity modulus is bounded above and below by positive constants, while the estimate (5.12.14) shows that, under a simple shear, the stress increases as the strain increases. We suppose also y ! (y) is a continuous mapping of R into R+ and 0 < (5.12.15) a4 (y) a5 for all y 2 R, y ! (y) is a continuous mapping of R into R and j (y)j (5.12.16) a6 for all y 2 R, where a4 , a5 , a6 are constants. Finally, we assume that K = (K1 K2 K3) 2 (L2 ( ))3 (5.12.17) 3 F = (F1 F2 F3 ) 2 (L2 (S2 )) (5.12.18)

418

Chapter 5. Optimization of deformable solids

(5.12.19) De ne an operator L mapping H into H and an element G 2 H by (L v h) = 3 (G h) =

Z

2 W21 ( ):
Z

( )"(v)"(h) dx + 2

g ( I (v ) )

3 X

Z X 3

v h2H
i=1

i j =1

eij (v)eij (h) dx (5.12.20)

Ki hi dx +

Z X 3

A generalized solution of the problem (5.12.1), (5.12.6) is de ned to be a function u satisfying u2 H (L u h) = (G h) h 2 H: (5.12.22) Let us show that, if a function u is a classical solution of the problem (5.12.1), (5.12.6), then it also satis es (5.12.22). By (5.12.4) and (5.12.6), using Green's formula, we have, for all h 2 H ,

S2 i=1

Fi hi ds + 3

( ) ( ) "(h) dx: (5.12.21)

Z X 3

i=1

Ki hi dx =
=

Z X 3

Z X 3

@ ij (u ) h dx i i j =1 @xj Fi hi ds ;

Z X 3

S2 i=1

i j =1

ij (u )"ij (h) dx:

(5.12.23)

Therefore, by virtue of (5.12.1), (5.12.3), and (5.12.10), we get (5.12.22). Caring out inverse actions, one easily concludes that a smooth solution of the problem (5.12.22) is a classical solution of (5.12.1), (5.12.6).

(in the topology generated by the product of the corresponding weak topologies ) into the space H (equipped with the strong topology ). To prove Theorem 5.12.1, we need the following lemma. Lemma 5.12.2 Let the conditions of Theorem 5.12.1 be satis ed. Then, L is a strictly monotone, coercive, and continuous mapping of H into H , more exactly, the following relations hold : (L v ; L w v ; w) c
Z
1 1 ("(v ; w))2 + (I 2 (v) ; I 2 (w))2 dx

Existence theorem Theorem 5.12.1 Let Conditions I, II be satis ed, and let (5.12.15){(5.12.19) hold. Then, there exists a unique solution of the problem (5.12.22). Moreover, the function (K F ) ! u is a continuous mapping of (L2 ( ))3 (L2 (S2 ))3 W21 ( )

v w2 H

2 W21 ( ) c = min(2a3 3a4)

(5.12.24)

5.12. Optimization of the residual stresses in an elastoplastic body

419

and the condition (L v ; L w v ; w) = 0 implies that v = w

(L v v) c1 kvk2 H

v2H

2 W21 ( ) c1 = min(2a1 3a4 )

(5.12.25)

and nally the condition v(n) ! v(0) in H yields L v(n) ! L v(0) in H :

Proof. By using (5.12.2), (5.12.5), (5.12.20), upon the estimate

3 X

i j =1

eij (v)eij (w)

1 2 I 1 (v)I 2 (w)

we get (L v ; L w v ; w) = 3 +2
Z h

( )("(v ; w))2 dx
3 X

g(I (v) )I (v) + g(I (w) )I (w) ; g(I (v) )

3 X

; g(I (w) )
3
Z Z

i j =1

eij (w)eij (v) dx

i j =1

eij (v)eij (w)

( )(e(v ; w))2 dx
1 1 2 2 g(I (v) )I 1 (v) ; g(I (w) )I 1 (w) (I 2 (v) ; I 2 (w)) dx:

+2

(5.12.26)

Therefore, by using (5.12.14) and (5.12.15), we arrive to (5.12.24). Let now (L v ; L w v ; w) = 0: Then, because of (5.12.24), I (v) = I (w) a.e. in , and (5.12.20), (5.12.8), (5.12.13), (5.12.15) imply that 0 = (L v ; L w v ; w) = 3 +2
Z Z

( )("(v ; w))2 dx

g(I (v) )I (v ; w) dx c2 kv ; wk2 H

c2 > 0

that is, v = w. The inequality (5.12.25) follows from (5.12.13) and (5.12.15). It remains to show that L is a continuous mapping of H into H . Evidently,

L = L(1) + L(2)

(5.12.27)

420

Chapter 5. Optimization of deformable solids

Z Z

where L(1) and L(2) are the mappings of H into H de ned by the formulas (L(1)v

h) = 3

( )"(v)"(h) dx

(5.12.28) (5.12.29)

(L(2)v h) = 2

g(I (v) )

3 X

i j =1

eij (v)eij (h) dx:

From (5.12.2) and (5.12.15), it follows that L(1) is a linear continuous mapping of H into H . Let us show that L(2) is a continuous mapping. Let v(n) ! v(0) in H and let us choose from fv(n) g a subsequence fv(m) g such that

I (v(m) ) ! I (v(0) )
Denote

a.e. in :

(5.12.30) (5.12.31)

I (m) = I (v(m) )

e(m) = eij (v(m) ) ij m = 0 1 2 :::

g(m) = g(I (v(m) ) )

Now, taking (5.12.13) into account, we get (L(2) v(m) ; L(2)v(0) h) = 2

Z
3 X

i j =1

g(m) (e(m) ; e(0) ) + (g(m) ; g(0) )e(0) eij (h) dx ij ij ij

1 3 2 X Z ; (m) (g ; g(0))e(0) 2 dx ij i j =1

2 a2 kv(m) ; v(0) kH +

khkH :
(5.12.32)

By using Conditions I, II, the relations (5.12.30), (5.12.31), and the Lebesgue theorem, we infer that
Z ; lim (g(m) ; g(0) )e(0) 2 dx = 0: ij m!1

Hence, from (5.12.32), we get

L(2) v(m) ! L(2) v(0)

in H :

(5.12.33)

From an arbitrary subsequence fv(k) g chosen from the sequence fv(n) g one can choose, in turn, a subsequence fv(m) g satisfying (5.12.33), and therefore the latter formula holds not only for the subsequence fv(m) g, but for the whole sequence fv(n) g, i.e., L(2) v(n) ! L(2) v(0) in H , concluding the proof of the lemma.

5.12. Optimization of the residual stresses in an elastoplastic body

421

Proof of Theorem 5.12.1. 1) The existence of a unique solution of the prob-

lem (5.12.22) follows from Lemma 5.12.2 and from known results on monotone operators (see, e.g., Lions (1969), Gajewski et al. (1974)). It remains to prove the continuity of the function (K F ) ! u in the above mentioned sense. Let fK ( ) F ( ) g1 be a sequence such that =1 weakly in (L2 ( ))3 weakly in (L2 (S2 ))3 ! 0 weakly in W21 ( ): (5.12.34) Let us choose a subsequence f n g such that strongly in L2 ( ) and a.e. in : (5.12.35) n! 0 Upon (5.12.34), from the imbedding theorem (Theorem 1.6.2), it follows that K (n) ! K (0) strongly in H F (n) ! F (0) strongly in H (5.12.36) By (5.12.15), (5.12.16), we have
Z

K ( ) ! K (0) F ( ) ! F (0)

( n ) ( n ) n "(h) dx ; =
Z

( 0 ) ( 0 ) 0 "(h) dx (5.12.37)

( n ; 0) ( n) ( n)

+ ( ( n ) ( n ) ; ( 0 ) ( 0 )) 0 "(h) dx (ck n ; 0 kL2( ) + n )khkH where

n = k( ( n ) ( n ) ; ( 0 ) ( 0 )) 0 kL2 ( ) :

By using (5.12.15), (5.12.16), (5.12.35), and the Lebesgue theorem, we conclude lim = 0: (5.12.38) n!1 n Put (G(n) h) =
Z X Z X Z 3 3 Ki(n)hi dx + Fi(n) hi ds + 3 S2 i=1 i=1

( n ) ( n ) n "(h) dx (5.12.39) (5.12.40)

n = 0 1 2 :::

h 2 H:

From (5.12.35){(5.12.38), we have lim kG(n) ; G(0) kH = 0: n!1

422

Chapter 5. Optimization of deformable solids

Let u(n) be a solution of the following problem:

u(n) 2 H

(L n u(n) h) = (G(n) h)
n!1

h2H

n = 0 1 2 : : : (5.12.41)
(5.12.42)

We have to show that lim ku(n) ; u(0)kH = 0: From (5.12.25) and (5.12.41) we infer

c1 ku(n)k2 (L n u(n) u(n) ) = (G(n) u(n)) kG(n) kH ku(n)kH : H

Therefore, (5.12.40) yields

ku(n) kH const kL n u(n) kH

u(m) ! u L m u(m) !
const

8n: 8n:

(5.12.43) (5.12.44)

Upon (5.12.13), (5.12.15), (5.12.20), and (5.12.43), we get 2) Because of (5.12.43) and (5.12.44), from the sequence fu(n)g one can subtract a subsequence fu(m)g such that weakly in H weakly in H : (5.12.45) (5.12.46)

Passing to the limit in (5.12.41) (with n = m), taking to notice (5.12.40) and (5.12.46), we conclude ( h) = (G(0) h) and so = G(0) : On the other hand, (5.12.40), (5.12.41), (5.12.45), and (5.12.47) imply (L m u(m) u(m) ) = (G(m) u(m) ) ! (G(0) u) = ( u): By (5.12.24), (L m u(m) ; L m v u(m) ; v) 0 (5.12.48) (5.12.49) (5.12.47)

h2H

v 2 H:

By using Conditions I, II, the relation (5.12.35), and the Lebesgue theorem, we get lim m!1 kg (I (v ) m )eij (v ) ; g (I (v ) 0 )eij (v )kL2 ( ) = 0: (5.12.50)

5.12. Optimization of the residual stresses in an elastoplastic body

423

Analogously, by using (5.12.15) and (5.12.35), we obtain

m!1

lim k ( m )"(v) ; ( 0 )"(v)kL2 ( ) = 0: lim (L m v v) = (L 0 v v) (m) lim m!1(L m v u ) = (L 0 v u):

m!1

(5.12.51)

From (5.12.20), (5.12.45), (5.12.50), and (5.12.51) (5.12.52)

By using (5.12.45), (5.12.46), (5.12.48), and (5.12.52), we pass to the limit in (5.12.49), which gives ( ; L 0 v u ; v) 0

v 2 H:

(5.12.53)

Setting in (5.12.53) v = u ; w, > 0, w 2 H , we have ( ; L 0 (u ; w ) w ) 0 : Letting tend to zero, we conclude from Lemma 5.12.2 that ( ; L 0 u w) 0

w 2 H:

Hence = L 0 u = G(0) , i.e., with the notations (5.12.41), u = u(0) . Since the problem (5.12.41) has a unique solution, it is easy to see that (5.12.45) holds for the whole sequence fu(n)g, i.e.,

u(n) ! u(0)
3) Let us show that
n!1

weakly in H:

(5.12.54) (5.12.55) (5.12.56)

lim ku(n) kH = ku(0)kH :

Let

Xn = (L n u(n) ; L 0 u(0) u(n) ; u(0)):

By (5.12.41),

Xn = (G(n) ; G(0) u(n) ; u(0) ):

Thus, (5.12.40) and (5.12.43) give
nlim Xn = 0: !1

(5.12.57) (5.12.58)

Setting the notations

g(k p) = g(I (v(k) ) p )

k p = 0 1 2:::

424

Chapter 5. Optimization of deformable solids

by using (5.12.20), (5.12.31), and (5.12.56), we get

Xn =
where

4 X (i) Xn i=1

(5.12.59)

X (1) = 3
n

Z Z Z

( n )("(u(n) ; u(0) ))2 dx

3 X (n n) (n) (g eij ; g(0 n) e(0) )(e(n) ; e(0) ) ij ij ij i j =1 3 X (0 n) (g ; g(0 0))e(0) (e(n) ; e(0) ) dx: ij ij ij i j =1

(2) Xn = 3 ( ( n ) ; ( 0 ))"(u(0) )"(u(n) ; u(0) ) dx (3) Xn = 2 (4) Xn = 2

dx

(5.12.60)

By using the Cauchy inequality, we get 1 (2) 3 jXn j

Z

( ( n ) ; ( 0 ))2 ("(u(0) ))2 dx

Z
1 2 ("(u(n) ; u(0)))2 dx :

1 2

(5.12.61)

Utilizing (5.12.15), (5.12.35), and the Lebesgue theorem, it is not hard to see that the rst multiplier on the right hand side of (5.12.61) tends to zero as n ! 1. The second multiplier in (5.12.61) is bounded for each n, by virtue of (5.12.54). Hence,
n!1
(2) lim jXn j = 0: (4) lim jXn j = 0:

(5.12.62)

Analogously, we establish that

n!1

Hence, (5.12.57), (5.12.59), and (5.12.62) imply

n!1
(1) (3) lim (Xn + Xn ) = 0:

This equality together with (5.12.24) gives

"(u(n) ) ! "(u(0) ) in L2 ( ) 1 1 I 2 (u(n) ) ! I 2 (u(0) ) in L2 ( )

(5.12.63)

5.12. Optimization of the residual stresses in an elastoplastic body

425

Since y ! kykL2( ) is a continuous mapping of L2( ) into R, (5.12.8) and (5.12.63) imply (5.12.55). Now, on account of (5.12.54) and (5.12.55), we have

u(n) ! u(0)

strongly in H:

(5.12.64)

It remains to prove that the whole sequence fu g strongly converges to u(0) in H . By using the above argument and the fact that the problem (5.12.22) has a unique solution, one sees that, from an arbitrary subsequence fu(k) g extracted from the sequence fu( )g, one can extract, in turn, a subsequence fu(n)g for which (5.12.64) holds. Hence, u( ) ! u(0) strongly in H , concluding the proof.

Continuity of the mappings

Denote by u(K F ) the element u of H solving the problem (5.12.22), and de ne functions

K F ! ij (u(K F ) )
where ij (u )'s are given by (5.12.1).

i j=1 2 3

(5.12.65)

Theorem 5.12.2 Let Conditions I, II and the relations (5.12.15), (5.12.16) hold. Proof. Let fK (n)g, fF (n)g, f ng be sequences such that K (n) ! K (0) weakly in (L2 ( ))3 F (n) ! F (0) weakly in (L2 (S2 ))3 weakly in W21 ( ): n! 0
Introduce the notation

Then, the functions (5.12.65) are continuous mappings of (L2 ( ))3 (L2 (S2 ))3 W21 ( ) (in the topology generated by the product of the corresponding weak topologies ) into L2 ( ) endowed with the strong topology.

(5.12.66) (5.12.67) (5.12.68) (5.12.69) (5.12.70)

u(n) = u(K (n) F (n) n ) u(n) ! u(0)

n = 0 1 2:::
in H:

From (5.12.66){(5.12.68) and Theorem 5.12.1, we infer Because of (5.12.68) and (5.12.70), from the sequence f n u(n) g one can subtract a subsequence f m u(m)g such that
m! 0 (m) ) ! I (u(0) ) I (u

strongly in L2 ( ) and a.e. in a.e. in :

(5.12.71) (5.12.72)

426

Chapter 5. Optimization of deformable solids

We have kg(I (u(m)) m )eij (u(m) ) ; g(I (u(0)) 0 )eij (u(0) )kL2 ( ) kg(I (u(m)) m )(eij (u(m) ) ; eij (u(0) ))kL2 ( ) ; + g(I (u(m) ) m ) ; g(I (u(0)) 0 ) eij (u(0) ) L2 ( ) :

(5.12.73)

By (5.12.13) and (5.12.70), the rst term on the right hand side of the inequality (5.12.73) tends to zero as m ! 1. From Conditions I, II and the relations (5.12.71) and (5.12.72), by virtue of the Lebesgue theorem, we get that the second term in (5.12.73) also tends to zero. So,

g(I (u(m) ) m )eij (u(m) ) ! g(I (u(0) ) 0 )eij (u(0) )

Analogously, we establish that

in L2 ( ):

(5.12.74)

( m )("(u(m) ) ; ( m ) m ) ! ( 0 )("(u(0) ) ; ( 0 ) 0 ) Hence, (5.12.1) and (5.12.74) give

ij (u(m) m ) ! ij (u(0) 0 )

in L2 ( ): (5.12.75)

in L2 ( )

i j = 1 2 3:

Analogously to the above, we verify that, from an arbitrary subsequence fu( ) g chosen from the sequence fu(n) n g, one can extract, in turn, a subsequence fu(m) m g for which (5.12.75) holds. Thus,
ij (u(n) n ) ! ij (u(0) 0 )

in L2 ( )

concluding the proof. Suppose that a nonlinear elatoplastic body is a ected by elds of forces K , F and a temperature pattern that satisfy the conditions (5.12.17){(5.12.19). The stresses ij (u ) and strains "ij (u ) in this body are de ned by the formulas (5.12.1) and (5.12.4), with u the solution of the problem (5.12.22). Let, further, there happen unloading and let K re, F re , re be the elds of forces and temperature pattern at the end of unloading. We assume that the unloading is complete, that is, K re = 0, F re = 0, re = 0. Then, in the body under consideration, there will appear residual re stresses ij , strains "re, and displacements (deformations) ure, which are given by ij the formulas (see Shevchenko (1970)) (p) re g (0 0) ij = g (0 ) ij (u ) ; ij "re = "ij (u) ; "(p) ij ij re = u ; u(p) ui (5.12.76) i i

5.12.2 Residual stresses and deformations

5.12. Optimization of the residual stresses in an elastoplastic body

(p) (p) (p) ij = (0)("(u ) ; (0) ) ij + 2g (0 0)eij (u ) "(p) = "ij (u(p) ) ij

427 (5.12.77) (5.12.78) (5.12.79) (5.12.80)

where the function u(p) is the solution of the problem

u(p) 2 H
(Lw h) = 3 (Gp h) = (G(1) h) = p
Z

(Lu(p) h) = (Gp h)
Z

h 2 H: eij (w)eij (h) dx

Here, we have used the following notations: (0)"(w)"(h) dx + 2 (G(i) h) p

g(0 0)

3 X

4 X

w h2H

i j =1

(G(4) h) = 3 (0) (0) p

3 g(0 0) X K h dx g(0 ) i=1 i i Z 3 X (0 @ (G(2) h) = ; (gg(0 0) 2 @g (0 ) ij (u ) @x hj dx p )) @ i i j =1 Z 3 g(0 X (G(3) h) = g(0 0) Fi hi ds p ) i=1 S2

i=1 Z

"(h) dx:

(5.12.81)

The relations (5.12.76){(5.12.81) follow from the rst theorem of unloading, which is derived under the supposition that, during the unloading, there are no additional plastic deformations (i.e., the unloading is going on elastically), and that the functions ( ) and g(0 ) are connected by the relation )(1 ( ) = 2g(01 ; 2 + ) where is the Poisson ratio|a constant independent of the temperature , 0 1 < 2. We suppose that, in addition to Conditions I, II, the function g satis es also the following one: ! g(0 ) is a continuously di erentiable mapping of R 9 > > = into R and (5.12.82) @g (0 ) a > > 2 R: 7 @

428

Chapter 5. Optimization of deformable solids

Lemma 5.12.3 Let Conditions I, II and the relations (5.12.15){(5.12.18), (5.12.82) hold. Suppose that 2 W22 ( ) and ij (u ) are de ned by the formula (5.12.1) in which u is the solution of the problem (5.12.22). Then, the element Gp de ned by the formula (5.12.81) belongs to the space H , and the function (K F ) ! Gp is a bilinear mapping of (L2 ( ))3 (L2 (S2 ))3 W22 ( ) (in the topology generated by the product of the corresponding weak topologies ) in the space H (endowed with the strong topology). Proof. Let fK (n) F (n) ng be a sequence such that K (n) ! K (0) weakly in (L2 ( ))3 (5.12.83) (n) ! F (0) 3 F weakly in (L2 (S2 )) (5.12.84) 2 ( ): weakly in W2 (5.12.85) n! 0 Then, by the imbedding theorem, we have strongly in C ( ): (5.12.86) n! 0 For any z 2 L2( ),
(Ki(n) ; Ki(0))z dx g(0 n ) Z + Ki(0) z g(01 ) ; g(01 ) dx : n 0 (5.12.87) Upon the Lebesgue theorem, by using Conditions I, II and the relation (5.12.86), we get Hence, (5.12.83) yields that the rst term on the right hand side of (5.12.87) tends to zero as n ! 1. Absolutely analogously, we conclude that so does the second term in (5.12.87). Thus,
Z

Ki(n) ; Ki(0) z dx g(0 n ) g(0 0)

! g(0z ) g(0 n ) 0

strongly in L2 ( ):

Fi(n) ! Fi(0) (5.12.89) g(0 n) g(0 0 ) weakly in L2(S2 ): Let us denote by Gpn and G(i) the functionals Gp and G(i) , respectively, pn p when K = K (n) , F = F (n) , = n , n = 0 1 2 : : : From (5.12.88) and Theorem 1.5.12 it follows that G(1) ! G(1) in H . pn p0

Analogously,

Ki(n) ! Ki(0) g(0 n) g(0 0)

weakly in L2( ):

(5.12.88)

5.12. Optimization of the residual stresses in an elastoplastic body

429

Let W stand for the set of functions from W21 ( ) vanishing on S1 , W is a 1 2 closed subspace of W21 ( ). Denote by H00 (S2 ) the space of the traces on S2 of the functions from W . This space equipped with the norm 1 k kH 2 (S ) = inf kvkW21 ( ) v 2 W v S2 =
00 2 1 2 is a Hilbert space, and the following imbedding takes place H00 (S2 ) H 2 (S2 ), see Lions and Magenes (1972). 1 1 1 2 2 2 Let (H00 (S2 )) be the dual space of H00 (S2 ). The space H00 (S2 ) is compactly imbedded into L21(S2 ) (see Remark 1.6.1), and therefore L2 (S2 ) is com2 pactly imbedded into (H00 (S2 )) (see Theorem 1.5.12). Hence, (5.12.89) gives that G(3) ! G(3) in H . pn p0 From (5.12.86), we conclude that G(4) ! G(4) in H . It remains to show that pn p0 1

@g @n @g @ 0 in L ( ): (5.12.91) ijn ij 0 10=7 2 @ (0 n ) @xi ! (g (0 0 ))2 @ (0 0 ) @xi (g(0 n )) Here, ijn stand for the stresses ij de ned by the formula (5.12.1) in which u is the solution of the problem (5.12.22) for K = K (n), F = F (n) , = n . By (5.12.83){(5.12.85) and Theorem 5.12.2, we have ijn ! ij0 in L2 ( ).
Hence, taking to notice Condition I and (5.12.13), (5.12.82), (5.12.86), we get

in H : (5.12.90) Since H is continuously imbedded into (L10=3 ( ))3 , in order to prove (5.12.90) it su ces to show that

G(2) ! G(2) pn p0

@g @g ijn ij 0 (g(0 n ))2 @ (0 n ) ! (g(0 0))2 @ (0 0 ) Upon (5.12.85) and the imbedding theorem, @ n!@ 0 @xi @xi

in L2 ( ):

(5.12.92) (5.12.93)

strongly in L5 ( ):

Since y0 y00 ! y0 y00 is a bilinear continuous mapping of L2 ( ) L5 ( ) into L10=7 ( ), (5.12.91) follows from (5.12.92) and (5.12.93). The lemma is proved. Theorem 5.12.3 Let Conditions I, II be satis ed, let the relation (5.12.15){ (5.12.18), (5.12.82) hold, and let 2 W22 ( ). Then, there exists a unique solution of the problem (5.12.79), and the function K F ! u(p) de ned by this solution is a continuous mapping of (L2 ( ))3 (L2 (S2 ))3 W22( ) (in the topology generated by the product of the corresponding weak topologies ) into the space H (endowed with the strong topology ). Proof. Evidently, the bilinear form w h ! (Lw h) de ned by (5.12.80) is symmetric and continuous on H H . By (5.12.5) and (5.12.8), this form is coercive on H . Now, the theorem follows from Lemma 5.12.3 and the Riesz theorem.

430

Chapter 5. Optimization of deformable solids

A stationary temperature pattern in a heated body is de ned as the solution of the heat equation

5.12.3 Temperature pattern in a medium

k
3 X @2

where k is the thermal conductivity, ' is the function of density of heat sources. We take the boundary condition in the form of the Newton law where is the unit outward normal to the surface, q the coe cient of the heat transfer, 0 the temperature of the environment. We suppose that

i=1

@x2 = ;' i

in

(5.12.94)

@ @ = ;q( ; 0 )

on S

(5.12.95)

k = const > 0

q = const > 0:

(5.12.96) (5.12.97) (5.12.98)

Letting q 0 = I , we write down the boundary condition (5.12.95) in the form

@ @ + q = I: Suppose S is a surface of the C 2 class and 1 ' 2 L2( ) I 2 H 2 (S ):

A generalized solution of the problem (5.12.94), (5.12.97) is de ned to be a function such that

2 W21 ( )
where

a( w) =
Z X 3

'w dx + k

I w ds

w 2 W21 ( ) u v 2 W21 ( ):

(5.12.99)

a(u v) = k

@u @v dx + qk Z uv ds S i=1 @xi @xi

(5.12.100)

It is not hard to see that a smooth generalized solution of the problem (5.12.94), (5.12.97) is a classical solution of this problem, and vise versa, a classical solution is a generalized one. Evidently, the bilinear form a(u v) is symmetric and continuous on W21 ( ) 1 ( ). From the Friedrichs inequality we conclude the coercivity of this form. W2 Therefore, the problem (5.12.99) has a unique solution. Thus, from known results on smoothness of generalized solutions of elliptic problems, see Agmon et al. (1959), Lions and Magenes (1972), we get the following

5.12. Optimization of the residual stresses in an elastoplastic body

431

Theorem 5.12.4 Let be a bounded domain in R3 with the boundary of the C 2 class, and let (5.12.96), (5.12.98) hold. Then, there exists a unique solution of the problem (5.12.94), (5.12.97), and the function ' I ! is a continuous mapping 1 of L2( ) H 2 (S ) into W22 ( ).
By virtue of Theorem 5.12.4, we can de ne a linear continuous operator A 2 mapping L2 ( ) H 1 (S ) into W22 ( ) such that A(' I ) = , where is the solution of the problem (5.12.94), (5.12.97). Therefore, for any xed h 2 (W22 ( )) 1, the mapping ' I ! (A(' I ) h) is a linear continuous functional on L2 ( 1 ) H 2 (S ). Hence, the conditions 'n ! ' weakly in L2 ( ), In ! I weakly in H 2 (S ) yield (A('n In ) h) ! (A(' I ) h): Thus, we have Corollary 5.12.1 Under the conditions of Theorem 5.12.4, the function ' I ! 2 is a continuous mapping of L2( ) H 1 (S ) (in the topology generated by the product 2 of the weak topologies of L2 ( ) and H 1 (S )) into the space W22 ( ) endowed with the weak topology. Taking to notice the above results, for a given power load K , F and a given thermal load ', I , one can calculate the residual stresses, strains, and displacements in the following way. From the solution of the problem (5.12.94), (5.12.97), one nds the function . Then, one calculates the function u from the solution of the problem (5.12.22), and by using the formula (5.12.1), one derives the stresses under the load ij (u ). Next, by (5.12.81), one nds the functional Gp , and from the solution of the problem (5.12.79) one gets the function u(p) . Finally, by using the formulas re (5.12.76){(5.12.78), one obtains ij , "re, ure. ij i We suppose that the functions K and F (volume and surface forces), as well as the functions ' and I (thermal load) are controls. Denote by t the control vector function t = (K F ' I ) 2 U , where U is the space of controls, which is taken in the form 1 U = (L2 ( ))3 (L2 (S2 ))3 L2 ( ) H 2 (S ): (5.12.101) re Evidently, the residual stresses ij , strains "re, and displacements ure de ned ij i by the formula (5.12.76) depend on an element t 2 U . This is why, in what follows, (re we denote them by ij t) , "(re t) , u(re t) . Let (re t) , "(re t) , u(re t) stand for the tensors ij i of residual stresses, of residual strains, and for the vector function of residual displacements, respectively. By virtue of the above results,
(re t) (re = f ij t) g 2 (L2 ( ))6 "(re t) = f"(re t) g 2 (L2 ( ))6 ij (re t) = fu(re t) g 2 (W 1 ( ))3 : u 2 i

5.12.4 Optimization problem

432

Chapter 5. Optimization of deformable solids

Notice that, because of the symmetricity of the tensors of stresses and strains, ( these are considered as elements of the space (L2 ( ))6 . Finally, by (t) = f ijt) g we denote the tensor of the stresses that appear during the loading. These stresses are given by the formula (5.12.1) via the solutions of the problems (5.12.22) and (5.12.94), (5.12.97). More exactly, from the problem (5.12.94), (5.12.97), one nds the function , which is used in the solution of the problem (5.12.22). Assume, next, that we are given functionals k such that

u ! k (t 0 " 00 u) is a continuous mapping9 > > > of U (L2 ( ))6 (L2 ( ))6 (L2 ( ))6 (W21 ( ))3 (in the> > = t
0

"

00

topology generated by the product of the weak topology of> > the space U and the strong topology of the space (L2 ( ))6 > > > 6 (L2 ( ))6 (W 1 ( ))3 ) into R, k = 0 1 : : : l. (L2 ( )) 2

(5.12.102)

De ne a set of admissible controls as follows: Uad = t j t = (K F ' I ) 2 U ktkU c k (t (re t) "(re t) (t) u(re t) ) ck k = 1 2 : : : lg (t) =
0 (t (re t) "(re t) (t) u(re t) ):

(5.12.103)

where c, ck are positive constants, and let a goal functional be of the form (5.12.104) (5.12.105) The optimization problem consists in nding a function y such that

y 2 Uad
Z X 3

(y) = t2U inf

ad

(t):

The functional (t) can be taken, for instance, in the form (t) =
i j =1
(re ( ij t) ; aij )2 dx

(5.12.106)

where aij are given elements of the space L2 ( ). Then, the problem (5.12.105) means that one must nd a control for which the distribution of the residual stresses is the mean square closest to a given one. The functionals k , k = 1 2 : : : l, can be taken so that the conditions
k (t (re t) "(re t) (t) u(re t) )

ck

will de ne restrictions on the residual stresses, strains, and displacements, as well as restrictions on strength. The latter ones are constructed for the tensor (t) , by using the strength criteria (see subsec. 5.1.2), in order that the deformable body do not fail by the load.

5.12. Optimization of the residual stresses in an elastoplastic body

433

(5.12.15), (5.12.16), (5.12.82), (5.12.96), (5.12.102){(5.12.104) hold, while the set Uad is nonempty. Then, there exists a solution of the problem (5.12.105).

C 2 class, let Conditions I, II be satis ed for a function g, and let the relations

Theorem 5.12.5 Let

be a bounded domain in R3 with the boundary of the

Proof. Let ftn = (K (n) F (n) 'n In )g1=1 be a minimizing sequence, i.e., n ftn g Uad (5.12.107)
inf nlim (tn ) = t2Uad (t): !1 (5.12.108) Upon (5.12.107), the sequence ftn g is bounded in U . Hence, we can subtract from it a subsequence ftm = (K (m) F (m) 'm Im )g1=1 such that tm ! t0 weakly in m U , where t0 = (K (0) F (0) '0 I0 ). So

K (m) ! K (0) F (m) ! F (0) 'm ! '0

Im ! I0

weakly in (L2 ( ))3 weakly in (L2 (S2 ))3 weakly in L2( ) 1 weakly in H 2 (S ):

(5.12.109) (5.12.110) (5.12.111) (5.12.112)

I = Im , m = 0 1 2 : : : From (5.12.111), (5.12.112), because of Corollary 5.12.1,

it follows that
m! 0

Denote by m the solution of the problem (5.12.94), (5.12.97) with ' = 'm , weakly in W22 ( ): in C ( ):

(5.12.113) (5.12.114)

Hence, by the imbedding theorem,

m! 0

Let, further, u(m) be the solution of the problem (5.12.22) with K = K (m) , ( F = F (m) , = m , and let (tm ) = f ijtm ) g be the stress tensor de ned by the formula (5.12.1) with u = u(m) , = m , m = 0 1 2 : : : By (5.12.109), (5.12.110), (5.12.113), and Theorems 5.12.1 and 5.12.2 ,

i j = 1 2 3: ij ! ij Let also u(p m) be the solution of the problem (5.12.79) with K = K (m) , ( ( F = F (m) , = m . Let ijp m) stand for the function ijp) de ned by (5.12.77) with u(p) = u(p m) and = m . By (5.12.109), (5.12.110), (5.12.113), and Theorem 5.12.3, we get

u(m) ! u(0)
(tm )

(t0 )

in H in L2 ( )

(5.12.115) (5.12.116)

u(p m) ! u(p 0)

in H

(5.12.117)

434 and therefore

Chapter 5. Optimization of deformable solids

(p m) p0 ij ! ij

in L2( )

i j = 1 2 3:

(5.12.118) (5.12.119) (5.12.120) (5.12.121)

By using the formulas (5.12.76) and the notations introduced, we get

g(0 0) (tm ) (p m) g(0 m) ij ; ij "(re tm ) = "ij (u(m) ) ; "ij (u(p m) ) ij u(re tm ) = u(m) ; u(p m)
ij
(re tm ) = (re tm ) ! (re t0 )

where "ij (u) is de ned by the formula (5.12.4). By (5.12.114), (5.12.116), (5.12.118), (5.12.119) and Conditions I, II, we have
ij ij

in L2 ( ): in L2 ( ) in H:

(5.12.122) (5.12.123) (5.12.124)

Upon (5.12.4), (5.12.115), (5.12.117), (5.12.120), (5.12.121),

"(re tm ) ! "(re t0 ) ij ij (re tm ) ! u(re t0 ) u t0 2 Uad

concluding the proof.

Taking to notice (5.12.102){(5.12.104), (5.12.107){(5.12.112), (5.12.116), and (5.12.122){(5.12.124), we pass to the limit, which gives (t0 ) = t2U inf
ad

(t)

ses, strains, and deformations reduces to a problem of control by the right hand sides in the class of nonlinear elliptic problems. For approximate solution of this problem, one can use Theorem 3.2.2, as well as approaches based on the extending of the set Uad (see, for instance, Section 2.4).

Remark 5.12.1 The problem (5.12.105) on the creation of optimal residual stres-

Chapter 6

Optimization problems for steady ows of viscous and nonlinear viscous uids
\ \How big and beatiful the world is!" said the toad. \But one must look round in it and not just stay in one spot all the time." And he hopped into the vegetable garden." | H. Ch. Andersen \The Toad" In this chapter, we will formulate and investigate various problems of optimization of ows of nonlinear viscous, non-Newtonian uids. The majority of real uids are non-Newtonian and nonlinear viscous. A special case of the uids under consideration is a viscous, Newtonian uid, so that all the results of this chapter remain true for it.

6.1 Problem on steady ow of a nonlinear viscous uid

6.1.1 Basic equations and assumptions
The constitutive equation of an incompressible, nonlinear viscous uid is the following, see Litvinov (1982a), i j = 1 : : : n: (6.1.1) ij (p v ) = ;p ij + 2'(I (v ))"ij (v ) Here, ij (p v) are the components of the stress tensor which depend on the functions of pressure p and velocity v = (v1 : : : vn ), n is the dimension of the domain 435

436

Chapter 6. Optimization problems for steady ows

:::

in which the ow is studied, ij is the Kronecker delta, "ij (v) are the components of the rate of strain tensor: @v "ij (v) = 1 @xi + @vj i j = 1 ::: n (6.1.2) 2 j @xi ' is the viscosity function depending on the second invariant of the rate of strain tensor I (v):

I (v) =

n X; i j =1

"ij (v) 2 :

(6.1.3)

In the case when ' = const > 0, the formula (6.1.1) de nes the constitutive equation of a viscous, Newtonian uid. We consider ows of a nonlinear viscous uid in a bounded domain Rn with a Lipschitz continuous boundary S . The equations of motion and incompressibility are de ned by @vi + v @vi ; @ ij (p v) = K in i = 1 ::: n (6.1.4)

@t

@xj

i @xj n X @v div v = @xi = 0 i=1 i

in :

(6.1.5)

Here and below, the summation over repeated indices is implied, is the density of the uid, Ki are the components of the vector function of volume forces K = (K1 : : : Kn). We consider steady, slow ows, and so in the equations (6.1.4) we @vi ignore the convection terms vj @xj . Thus, (6.1.1) and (6.1.4) lead to the following equations

@p ; 2 @ ;'(I (v))" (v) = K ij i @xi @xj

in

i = 1 : : : n:

(6.1.6)

Let S1 , S2 , S3 be open subsets of S such that S = S 1 S 2 S 3 , Si \ Sj = ? for i 6= j . We consider the mixed boundary conditions prescribing zero velocities on S1 , the surface forces F = (F1 : : : Fn ) on S2 , and the condition of ltration on S3 , see Fig. 6.1.1, i.e.,

v S1 = 0 ; p ij + 2'(I (v))"ij (v) j S2 = Fi i = 1 ::: n ^i i S3 = ; (v2 )v S3 ; p ij + 2'(I (v))"ij (v) j i S3 ;F (vi ; v i ) S3 = 0 i = 1 : : : n:

(6.1.7) (6.1.8) (6.1.9) (6.1.10)

Here, j are the components of the unit outward normal = ( 1 : : : n ) to S , v = vi i , is the function of ltration, depending on v2 and a point s 2 S3 . The function takes the same values when the uid is owing inwards or outwards the

6.1. Problem on steady ow of a nonlinear viscous uid

437

^ ^ ^ domain , so we write (v2 s), F = (F1 : : : Fn ) is the function of surface forces 0 acting on the external surface of the ltration layer S3 , see Fig. 6.1.1. We suppose ^ that the thickness of the ltration layer is small and carry the function F from 0 S3 to S3 . If the uid ows out into the void, or even into the air, we can take ^ F = 0. The condition (6.1.10) means that the velocity vector is normal to S3 . The case when ( s) = 0 ( s) 2 R+ S3 where R+ = j 2 R 0 corresponds to the situation when there is no ltration layer on S3 and the normal ^ component of the surface forces is equal to Fi i . The case when ( s) = 1 for each ( s) 2 R+ S3 corresponds to the solid wall and the adhesion on S3 . Indeed, in that case, by (6.1.9) and (6.1.10), we have v = 0 on S3 . In particular, S3 may be an empty set, then the conditions (6.1.9), (6.1.10) are omitted.
S1
' S3

S2

S3
d

$ F

S1

Figure 6.1.1: Domain and the ltration layer on S3 We suppose that

' is a function continuously di erentiable in R+ , and there9 = exist positive constants a1 : : : a4 such that, for an arbitrary 2 R+ , the following inequalities hold: a1 '( ) a2

(6.1.11) (6.1.12) (6.1.13) (6.1.14)

'( ) + 2 d' ( ) a3 d d' ( ) a : 4 d

The physical meaning of (6.1.12) is that the function of viscosity must be bounded by two positive constants. The inequality (6.1.13) means that, in the case of a simple shearing ow, the shearing stress increases as the shearing rate

438

Chapter 6. Optimization problems for steady ows

:::

increases, see Litvinov (1982a). The inequality (6.1.14) is a restriction on the increase of d' for large . All these assumptions are physically natural, see Litvinov d (1982a). Concerning the function , we take assumptions similar to the ones concerning ', namely, at each point s 2 S3 , the function ( s): ! ( s) is> > > continuously di erentiable in R+ , and there exist positive= constants b1, b2 such that, for an arbitrary ( s) 2 R+ S3 ,> > > the following inequalities hold 0 ( s) b1
9

(6.1.15) (6.1.16) (6.1.17) (6.1.18)

@ ( s) + 2 @ ( s) 0 @ ( s) b : 2 @

From the physical point of view, the estimate (6.1.17) means that the normal component of the surface force acting on the ltration layer does not decrease as the normal component of the velocity increases.

6.1.2 Formulation of the problem

We de ne spaces X and V as follows

X = u j u 2 H 1 ( )n u S1 = 0 (u ; u ) S3 = 0 V = u j u 2 X div u = 0
Z

(6.1.19) (6.1.20)

where is the unit outward normal to S , u = ui i . In virtue of Korn's inequality, the expression

kukX =
Z

I (u) dx

1 2

(6.1.21)

de nes norms in X and V that are equivalent to the norm of H 1 ( )n , X and V being Hilbert spaces with the scalar product (u h)X =

"ij (u)"ij (h) dx:

We de ne a function of surface forces F acting on S2 S3 by

F (s) = F (s) if s 2 S2 ^ F (s) if s 2 S3 :

6.1. Problem on steady ow of a nonlinear viscous uid

439 (6.1.22)

We suppose that If K 2 L2 ( )n , F 2 L2(S2 S3 )n , then (K h) =

Z

K F 2X :
Z

Ki hi dx
Z

(F h) =

Let us de ne operators L : X ! X and B 2 L(X L2 ( )) as follows (L(u) h) = 2 where

S2

Fi hi ds +
Z

S3

^ Fi hi ds

h 2 X:
(6.1.23) (6.1.24)

'(I (u))"ij (u)"ij (h) dx + Bu = div u

S3

(juj2 )ui hi ds

ju j =

By B we denote the operator adjoint of B , B 2 L(L2 ( ) X ) and (Bu ) = (u B ) u 2 X 2 L2 ( ): From (6.1.19) it follows that
Z

n 2 X 2 1 ui : i=1

We consider the problem: Find a pair of functions v, p satisfying (v p) 2 X L2 ( ) (6.1.26) (L(v) h) ; (B p h) = (K + F h) h2X (6.1.27) (Bv q) = 0 q 2 L2 ( ): (6.1.28) By using Green's formula, it may be shown (see Litvinov (1982a)) that, if v, p is a solution to the problem (6.1.26){(6.1.28), then v, p is a solution to the problem (6.1.6){(6.1.10) in the distribution sense. On the contrary, if v, p is a solution to (6.1.6){(6.1.10) that satis es (6.1.26), then v, p is a solution to (6.1.26){(6.1.28). In what follows, we will use the following Lemma 6.1.1 Let be a bounded domain in Rn , n = 2 3, with a Lipschitz continuous boundary S , and let S1 , S2 , S3 be open subsets of S such that S = S 1 S 2 S 3 , Si \ Sj = ? for i 6= j , S1 , S2 being nonempty. Let spaces X and V be de ned by (6.1.19), (6.1.20), and let an operator B 2 L(X L2 ( )) be de ned by (6.1.24). Then, there exists a positive constant 1 such that : (6.1.29) inf sup (Bv )
2L2 (

S3

(juj2 )ui hi ds =

S3

(u2 )u h ds

u h 2 X:

(6.1.25)

) v 2X

kvkX k kL2(

440

Chapter 6. Optimization problems for steady ows

:::

The operator B is an isomorphism of V ? onto L2( ), where V ? is the orthogonal complement to V in X , and the operator B is an isomorphism of L2 ( ) onto V 0 , where V 0 = f j f 2 X (f u) = 0 u 2 V (6.1.30) and moreover, 1 kB ;1 k ? (6.1.31)
L(L2 (

)V ) ))

k(B ) kL(V 0 L2 (
;1

1:
1

(6.1.32)

In order to prove Lemma 6.1.1, we use the following result, see Ladyzhenskaya and Solonnikov (1976), Girault and Raviart (1981). Lemma 6.1.2 Let be a bounded domain in Rn , n = 2 3, withR a Lipschitz continuous boundary. Then, for an arbitrary h 2 L2( ) such that h dx = 0, 1 there exists a function u 2 H0 ( )n for which div u = h. Proof of Lemma 6.1.1. First, let us show that the operator B is an isomorphism of V ? onto L2( ). By Lemma 6.1.2 and Banach's theorem, it su ces to show that there exists a function w such that w2X div w = 1: (6.1.33) Take a function g such that

g 2 H 1 ( )n

supp g

S2

Then, g 2 X and by Green's formula and (6.1.34), we obtain Z mes div g ; 1 dx = 0

S2

gi i ds = a 6= 0:

(6.1.34)

(6.1.35)

and for the function h = mes div g ; 1, we have a

h 2 L2 ( )

h dx = 0:

(6.1.36)

1 By (6.1.36) and Lemma 6.1.2, there exists a function u 2 H0 ( )n such that mes g ; u satis es (6.1.33). div u = h, and so the function w = a Therefore, the operator B is an isomorphism of V ? onto L2( ), and there exists a positive constant 1 such that (6.1.31) holds. The space V 0 can be identied isometrically with (V ? ) , and so, taking to notice (6.1.31) and the following equalities: (B ;1 ) = (B );1

6.1. Problem on steady ow of a nonlinear viscous uid

441

kB ;1 kL(L2(
we obtain (6.1.32). Hence,

) V ? ) = k(B

;1

) )kL(V 0 L2 (

))

kB kV 0 = sup? (vkvB ) = sup (vkvB ) kX v2X kX v 2V 2 L2 ( ) 1 k kL2 ( )

thus (6.1.29) holds and the lemma is proven. Let fXk g1 , fMk g1 be sequences of nite-dimensional subspaces of X k=1 k=1 and L2 ( ), respectively, such that
k!1 y2Mk

lim inf kw ; ykL2( ) = 0

Z

k!1 z2Xk

lim inf ku ; z kX = 0

u2X w 2 L2 ( ) u 2 Xk

(6.1.37) (6.1.38)

De ne operators Bk 2 L(Xk Mk ) as follows: (Bk u ) = div u dx

2 Mk

(6.1.39)

and let Bk 2 L(Mk Xk ) be the operator adjoint of Bk : (Bk u ) = (u Bk ) u 2 Xk 2 Mk : We introduce spaces Vk and Vk0 by

Vk = u j u 2 Xk (Bk u ) = 0 2 Mk : Vk0 = q j q 2 Xk (q u) = 0 u 2 Vk :

(6.1.40) (6.1.41)

spaces of X and L2 ( ), respectively, and let there exist a positive constant such that inf sup (Bk u ) 8k: (6.1.42)
2Mk

Lemma 6.1.3 Let fXk g1 , fMk g1 be sequences of nite-dimensional subk=1 k=1

u2Xk

kukX k kL2(

Then, the operator Bk is an isomorphism of Mk onto Vk0 , and the operator Bk is an isomorphism of Vk? onto Mk , where Vk? is the orthogonal complement to Vk in Xk , and (6.1.43) k(Bk );1 kL(Vk0 Mk ) 1 8k ; kBk 1 kL(Mk8 Vk? ) 1 8k (6.1.44)

442

Chapter 6. Optimization problems for steady ows

:::

Proof. It follows from (6.1.42) that

B sup (u ukk ) u2Xk k X
Therefore,

k kL2( k kL2(
)

2 Mk :

2 Mk (6.1.45) and Bk is an isomorphism of Mk onto its range R(Bk ). By the closed range theorem, R(Bk ) = Vk0 . The estimate (6.1.43) follows rom (6.1.45). Identifying
For the case when the velocity function satis es the zero boundary condition on the whole boundary, the subspaces Xk , Mk satisfying (6.1.37), (6.1.38), and (6.1.42) are contained in Girault and Raviart (1981), Gunzburger (1986). The extensions of these subspaces corresponding to the boundary conditions (6.1.6), (6.1.9) also satisfy (6.1.37), (6.1.38), and (6.1.42).

kBk kXk

Vk0 with (Vk? ) gives (6.1.44). The lemma is proved.

Lemma 6.1.4 Suppose the conditions (6.1.11){(6.1.18) are satis ed and an operator L is de ned by (6.1.23). Then, the following estimates hold :

(L(u) ; L(w) u ; w) kL(u) ; L(w)kX

2 1 ku ; wkX 2 ku ; wkX

u w2X u w2X

(6.1.46) (6.1.47)

where 1 , 2 are positive constants. Proof. Let u, w be arbitrary xed elements of X and let

h = u ; w:
We introduce a function as follows: (t) = 2
Z

(6.1.48)

'(I (w + th))"ij (w + th)"ij (e) dx

+
Z

S3

((w + th)2 )(w + th) e ds (0) = (L(w) e):

t 2 0 1] e 2 X:
(6.1.49)

By (6.1.23) and (6.1.25), we have (1) = (L(u) e) By using the theorem on the di erentiability of a function represented as an integral, see, e.g., Schwartz (1967), we conclude that is di ererentiable at any point t 2 (0 1). Thus, there exists 2 (0 1) such that (1) = (0) + d ( ) dt (6.1.50)

6.1. Problem on steady ow of a nonlinear viscous uid

443

where

d ( ) = 2 Z '(I (w + h))" (h)" (e) ij ij dt + 2 d' (I (w + h))"km (w + h)"ij (w + h)"km (h)"ij (e) dx Z h d i @ + ((w + h)2 )h e + 2 @ ((w + h)2 )(w + h)2 h e ds: (6.1.51)
S3

Now, by (6.1.11), (6.1.12), (6.1.14){(6.1.16), (6.1.18), (6.1.48){(6.1.51), we get (6.1.47). De ne functions g1 , g2 as follows:
8 > d' ( ) if d' ( ) < 0 < d g1 ( ) = > d :0 if d' ( ) 0 d 8 > @ ( s) if @ ( s) < 0 < @ g2( s) = > @ @ ( s) 0 :0 if

2 R+
( s) 2 R+ S3 :

(6.1.52) (6.1.53)

Taking e = h in (6.1.51) and applying (6.1.12), (6.1.13), (6.1.16), (6.1.17), and the estimate ; "ij (w + h)"ij (h) 2 I (w + h)I (h) we get

d ( ) = 2 Z '(I (w + h))I (h) + 2 d' (I (w + h))(" (w + h)" (h))2 dx ij ij dt d Z @ + ((w + h)2 )h2 + 2 @ ((w + h)2 )(w + h)2 h2 ds
2 +
Z S3 Z

'(I (w + h)) + 2g1(I (w + h))I (w + h) I (h) dx

((w + h)2 ) + 2g2 ((w + h)2 )(w + h)2 h2 ds

(6.1.54) where 1 is a positive constant. Now, (6.1.54) together with (6.1.48), (6.1.49) for e = h, and (6.1.50) gives (6.1.46), which completes the proof.

2 1 khkX

S3

6.1.3 Existence theorem

Consider another formulation of our problem: Find a function v satisfying v2V (6.1.55)

444

Chapter 6. Optimization problems for steady ows

:::

(L(v) h) = (K + F h)

h 2 V:

(6.1.56)

Clearly, if v, p is a solution of the problem (6.1.26){(6.1.28), then v is a solution of the problem (6.1.55), (6.1.56). On the contrary, if v is a solution of (6.1.55), (6.1.56), then L(v) ; K ; F 2 V 0 and due to Lemma 6.1.1, there exists a unique function p 2 L2( ) such that

L(v) ; K ; F = B p:
Therefore, the problems (6.1.26){(6.1.28) and (6.1.55), (6.1.56) are equivalent. We search for an approximate solution vk , pk of the problem (6.1.26){(6.1.28) in the form (vk pk ) 2 Xk Mk (L(vk ) h) ; (Bk pk h) = (K + F h) h 2 Xk (Bk vk q) = 0 q 2 Mk : (6.1.57) (6.1.58) (6.1.59)

that (6.1.37), (6.1.38), and (6.1.42) hold and Xk Xk+1 , Mk Mk+1 for each k. Then, for any k, there exists a unique solution vk , pk of the problem (6.1.57){ (6.1.59) and vk ! v in X , pk ! p in L2 ( ).

Theorem 6.1.1 Suppose the conditions (6.1.11){(6.1.18), (6.1.22) are satis ed. Then, there exists a unique solution v, p of the problem (6.1.26){(6.1.28). Let also fXk g, fMk g be sequences of nite-dimensional subspaces of X and L2( ) such

Proof. Due to the theory of monotone operators, see Gajewski et al. (1974),

Vainberg (1972), Varga (1971), the existence of a unique solution of the problem (6.1.55), (6.1.56) (and hence of the problem (6.1.26){(6.1.28)) follows from Lemma 6.1.4. We will prove the convergence of the approximate solutions vk , pk . The existence and uniqueness for the problem (6.1.26){(6.1.28) will also follow from this proof. 1. From (6.1.40), (6.1.57){(6.1.59), we get that the function vk is a solution of the problem

vk 2 Vk

(L(vk ) h) = (K + F h)

h 2 Vk :
0

(6.1.60) (6.1.61)

By (6.1.12), (6.1.16), (6.1.21){(6.1.23), we obtain, for an arbitrary u 2 Vk , (L(u) u) ; (K + F u) 2a1 kuk2 ; kK + F kX kukX X ;1 kK + F k : if kukX r = (2a1 ) X

So, because of a corollary of Brouwer's theorem, see Gajewski et al. (1974), Lions (1969), there exists a solution of the problem (6.1.60) and

kvk kX r

kL(vk )kX

(6.1.62)

6.1. Problem on steady ow of a nonlinear viscous uid

445

the second estimate in (6.1.62) follows from (6.1.12), (6.1.16), and the embedding theorem. For an arbitrary f 2 X , we denote by Gf the restriction of f to Xk , then Gf 2 Xk , and by virtue of (6.1.41), (6.1.60), we conclude

G(L(vk ) ; K ; F ) 2 Vk0 :
Thus, by Lemma 6.1.3, there exists a unique pk 2 Mk such that

Bk pk = G(L(vk ) ; K ; F )
and also vk , pk is a solution of the problem (6.1.57){(6.1.59) and

(6.1.63) (6.1.64) (6.1.65) (6.1.66) (6.1.67)

kpk kL2(

c1 :

By (6.1.62), (6.1.64), we can extract a subsequence fvm pmg such that

vm ! v0 weakly in X pm ! p0 weakly in L2( ) L(vm ) ! weakly in X :

Let m0 be a xed positive integer, and let h 2 Xm0 , q 2 Mm0 . By (6.1.65){ (6.1.67), we pass to the limit in (6.1.58), (6.1.59), with k replaced by m, which gives ( ; B p0 h) = (K + F h)
Z

(div v0 )q dx = 0

h 2 Xm0 q 2 Mm0 :

(6.1.68) (6.1.69)

Since m0 is an arbitrary positive integer, we infer from (6.1.37), (6.1.38), (6.1.68), and (6.1.69) that

; B p0 = K + F
div v0 = 0: From Lemma 6.1.4, we get (L(vm ) ; L(u) vm ; u) 0 u 2 X 8m: Since (Bm pm vm ) = 0, we get from (6.1.58), (6.1.65) that (L(vm ) vm ) = (K + F vm ) ! (K + F v0 ): The relations (6.1.67), (6.1.70) give lim m!1(L(vm ) u) ; (B p0 u) = (K + F u)

(6.1.70) (6.1.71) (6.1.72) (6.1.73) (6.1.74)

u 2 X:

446

Chapter 6. Optimization problems for steady ows

:::

By (6.1.73), (6.1.74), we pass to the limit in (6.1.72), and taking (6.1.71) into account, we get (K + F ; L(u) + B p0 v0 ; u) 0

u 2 X: h 2 X:

(6.1.75)

Take here u = v0 ; h, > 0, h 2 X , and let tend to zero. Then, due to the continuity of the mapping L (see (6.1.47)), we obtain (K + F ; L(v0 ) + B p0 h) 0 Therefore, the pair v = v0 , p = p0 is a solution of the problem (6.1.26){(6.1.28). 2. By using (6.1.73), (6.1.74), it is easy to see that (L(vm ) ; L(v0 ) vm ; v0 ) ! 0 and so, by (6.1.46), we get

vm ! v0
Let us show that

strongly in X: strongly in L2 ( ):

(6.1.76) (6.1.77)

pm ! p0

It follows from (6.1.27) and (6.1.58) that (L(v0 ) h) ; (B p0 h) = (K + F h) (L(vm ) h) ; (B pm h) = (K + F h) Thus, (B (pm ; ) h) = (L(vm ) ; L(v0 ) h) + (B (p0 ; ) h)

h 2 Xm h 2 Xm :
(6.1.78)

h 2 Xm

2 Mm :

The estimate (6.1.42) together with (6.1.78) yields kp ; k sup (B (pm ; ) h)

m L2 ( ) h2Xm
;1

khkX kL(vm) ; L(v0 )kX + ckp0 ; kL2 ( 2 Mm c = const > 0:

(6.1.79)
)

Hence,

kp0 ; pm kL2(

kp0 ; kL2( ) + kpm ; kL2( ) ;1 kL(vm ) ; L(v0 )kX + (c + 1) 2Mm kp0 ; kL2 ( ) : inf

(6.1.80)

6.2. Theorem on continuity

447

By (6.1.47) and (6.1.76), we get L(vm ) ! L(v0 ) in X , and (6.1.77) follows from (6.1.80) and (6.1.38). The function v0 is a solution of the problem (6.1.55), (6.1.56). Let v(1) , v(2) be two solutions of this problem, then (L(v(1) ) ; L(v(2) ) v(1) ; v(2) ) = 0 and by (6.1.46), we get v(1) = v(2) . Hence, the solution of (6.1.55), (6.1.56) is unique, and by Lemma 6.1.1 there exists a unique p0 such that the pair v = v0 , p = p0 is the unique solution of (6.1.26){(6.1.28). From the uniqueness of the solution, we infer that vk ! v0 in X and pk ! p0 in L2 ( ). The theorem is proved.

Remark 6.1.1 It is not hard to see that Theorem 6.1.1 remains true in the case
when S3 is an empty set. In this case, and (L(u) h) = 2

X = u j u 2 H 1 ( )n u S1 = 0
Z

(6.1.81) (6.1.82)

'(I (u))"ij (u)"ij (h) dx:

6.2 Theorem on continuity

It is obvious that the operator L is de ned by the functions of viscosity ' and ltration . So, we denote it by L' , and by (6.1.23), (6.1.25) we have (L' (u) h) = 2
Z

'(I (u))"ij (u)"ij (h) dx + u h 2 X:

S3

(u2 )u h ds

(6.2.1)

The solution of the problem (6.1.26){(6.1.28) depends on the loading K , F and ', . Therefore, we denote it by v(K F ' ), p(K F ' ). We have (v(K F ' ) p(K F ' )) 2 X L2( ) (L' (v(K F ' )) h) ; (B p(K F ' ) h) = (K + F h) (Bv(K F ' ) q) = 0 q 2 L2( ): We introduce a set B1 as follows (6.2.2) h 2 X (6.2.3) (6.2.4)

B1 = ' j ' 2 C 1 (R+ ) ' satis es the conditions (6.1.12){(6.1.14)

where a1 : : : a4 are positive constants :

(6.2.5)

448

Chapter 6. Optimization problems for steady ows

:::

C 1 (R+ ) being de ned by

We equip the set B1 with the topology generated by that of C 1 (R+ ), the norm of

k'kC 1(R+) = sup j'( )j + d' ( ) : d

2R+

(6.2.6)

We denote by L1(S3 C 1 (R+ )) the space of functions : S3 ! C 1 (R+ ) such that the function s ! k ( s)kC 1 (R+) is measurable and

k kL1(S3 C 1 (R+)) = sup ess k ( s)kC 1 (R+) < 1: s2S3

We introduce a set B2 by

(6.2.7)

B2 =

j 2 L1(S3 C 1 (R+ ))

satis es the conditions (6.1.16){(6.1.18) where b1 , b2 are positive constants : (6.2.8)

The set B2 is equipped with the topology generated by that of L1 (S3 C 1 (R+ )). By Theorem 6.1.1, the relations (6.2.2){(6.2.4) de ne the following function G:

B1 B2 3 (K F ' ) ! G(K F ' )

= (v(K F ' ) p(K F ' )) 2 V

L2 ( ): (6.2.9)

Theorem 6.2.1 The function G determined by (6.2.9), (6.2.2){(6.2.4) is a continuous mapping of X X B1 B2 into V L2( ). Proof. 1) Let f Km Fm 'm m g1=1 be a sequence such that m Km ! K0 in X (6.2.10) Fm ! F0 in X (6.2.11) 'm ! '0 in B1 (6.2.12) in B2 (6.2.13) m! 0
We introduce the notations vm = v(Km Fm 'm m ) pm = p(Km Fm 'm m )

m = 0 1 2 :::

(6.2.14)

where v(Km Fm 'm m ), p(Km Fm 'm m ) is the solution of the problem (6.2.2){(6.2.4) for K = Km , F = Fm , ' = 'm , = m . Then, the function vm is the solution of the following problem: (L'm
m (vm )

h) = (Km + Fm h)

h 2 V m = 0 1 2 ::: :

(6.2.15)

6.2. Theorem on continuity

449

We take here h = vm ; v0 and subtract (6.2.15) for m = 0 from (6.2.15). Then, (L'm

gm = (Km ; K0 + Fm ; F0 vm ; v0 ):

m (vm ) ; L'0 0 (v0 ) vm ; v0 ) = gm

(6.2.16) (6.2.17) (6.2.18)

Due to (6.1.12), (6.1.16), (6.2.10){(6.2.13), we have

kvm kX C
and so

8m

gm ! 0:
By (6.2.1) and (6.2.16)
4 X

i=1

Aim = gm

(6.2.19)

where

A1m = 2 A2m = 2 A3m = A4m =

Z

Z Z

'm (I (vm ))"ij (vm ) ; 'm (I (v0 ))"ij (v0 ) "ij (vm ; v0 ) dx 'm (I (v0 ))"ij (v0 ) ; '0 (I (v0 ))"ij (v0 ) "ij (vm ; v0 ) dx
2 2 m (vm )vm ; m (v0 )v0 (vm ; v0 ) ds 2 m (v0 2 )v0 ; 0 (v0 )v0 (vm ; v0 ) ds

(6.2.20)

ZS3

S3

Lemma 6.1.4 yields

A1m + A3m
We have

2 1 kvm ; v0 kX :

(6.2.21)

A2m 2 m A4m
m

"ij (v0 )"ij (vm ; v0 ) dx v0 (vm ; v0 ) ds

m = sup j'm (t) ; '0 (t)j t2R+ m = sup ess sup j m (t s) ; 0 (t s)j: s2S3 t2R+

S3

(6.2.22)

Thus, by (6.2.12), (6.2.13), (6.2.17), and (6.2.22), we get A2m ! 0, A4m ! 0, and (6.2.18), (6.2.19), (6.2.21) give

vm ! v0

in X , i.e., in V .

(6.2.23)

450

Chapter 6. Optimization problems for steady ows

:::

2) By (6.2.3) and (6.2.14), we have B (pm ; p0 ) = L'm m (vm ) ; L'0 0 (v0 ) ; Km + K0 ; Fm + F0 Let us show that From (6.2.1), we get (L'm 2
Z X n

in X . (6.2.24) (6.2.25)

L'm: m (vm ) ! L'0 0 (v0 ) h)

in X :

m (vm ) ; L'0 0 (v0 )

+
Z X n

i j =1 nZ S3

'm (I (vm ))"ij (vm ) ; '0 (I (v0 ))"ij (v0 ) 2 dx khkX

2 m (vm 2 )vm ; 0 (v0 )v0 2

1 2

ds 2 kh kL2 (S3 ) : (6.2.26)

1 2

o1

From the triangle inequality, it follows that

i j =1

'm (I (vm ))"ij (vm ) ; '0 (I (v0 ))"ij (v0 ) dx

B1m + B2m + B3m

(6.2.27)

where

B1m = B2m = B3m =

Z X n

i j =1 Z X n i j =1 Z X n i j =1

'm (I (vm ))"ij (vm ; v0 ) dx

1 2 1 2 1 2

('m (I (vm )) ; '0 (I (vm )))"ij (v0 ) dx

2

(6.2.28)

('0 (I (vm )) ; '0 (I (v0 )))"ij (v0 ) dx :

B1m a2 kvm ; v0 kX : (6.2.23) implies B1m ! 0, (6.2.12) gives B2m ! 0, and by the Lebesgue theorem, B3m ! 0.
Analogously, we get
Z

By (6.1.12) and (6.2.12), we have

and (6.2.25) follows from (6.2.26). Finally, (6.2.10), (6.2.11), (6.2.24), (6.2.25), and Lemma 6.1.1 give pm ! p0 in L2 ( ).

S3

2 m (vm

2 )vm ; 0 (v0 )v0

ds ! 0

as m ! 1

(6.2.29)

6.3. Continuity with respect to the shape of the domain

451

Let be a bounded domain in Rn , n = 2 3, with a Lipschitz continuous boundary S . Let M be a topological space, and to each q 2 M we assign a domain q Rn with a boundary Sq and a di eomorphism Pq of the C 1 class that maps q onto : Pq 2 C 1 ( q ) Pq;1 2 C 1 ( q ) (6.3.1) ;1 the di eomorphism inverse of P . For each q 2 M, consider the following Pq q problem on ow of a nonlinear viscous uid: Find a pair of functions vq , pq satisfying

6.3.1 Formulation of the problem

6.3 Continuity with respect to the shape of the domain

@pq ; 2 @ ;'(I (v ))" (v ) = K in q , i = 1 : : : n, (6.3.2) q ij q qi @xi @xj div vq = 0 in q (6.3.3) vq Sq1 = 0 (6.3.4) ; pq ij + 2'(I (vq ))"ij (vq ) qj Sq2 = Fqi i = 1 : : : n: (6.3.5) Here, Sq1 and Sq2 are open subsets of Sq such that Sq = S q1 S q2 , Sq1 \ Sq2 = ?, qj are the components of the unit outward normal q = ( q1 : : : qn ) to Sq , Kq = (Kq1 : : : Kqn ), Fq = (Fq1 : : : Fqn ) are functions of volume and surface forces given in q and Sq2 , respectively. So, to each q 2 M we assign the functions Kq and Fq . Let S1 , S2 be open subsets of S |the boundary of |for which S = S 1 S 2 , S1 \ S2 = ?.
We suppose that Pq is a bijection of Sq1 onto S1 and of Sq2 onto S2 for each (6.3.6) q 2 M. Speci cally, Sq2 and Fq may be xed, i.e., independent of q. De ne the following spaces Xq = u j u 2 H 1 ( q )n u Sq1 = 0 (6.3.7) Vq = u j u 2 Xq div u = 0 (6.3.8) 1 ( )n u = 0 : X= uju2H (6.3.9) S1 In Xq and Vq , the norm is de ned by (6.1.21) with = q , in X it is still de ned by (6.1.21). We suppose the function ' satis es the conditions (6.1.11){(6.1.14). For each q 2 M, de ne operators Lq : Xq ! Xq and Bq 2 L(Xq L2( q ) ) by (Lq (u) h) = 2
Z
q

'(I (u))"ij (u)"ij (h) dx

u h 2 Xq

(6.3.10)

452

Chapter 6. Optimization problems for steady ows

:::

(Bq u ) = We assume that

Z
q

(div u) dx

u 2 Xq

2 L2 ( q ):

(6.3.11)

Kq Fq 2 Xq
(vq pq ) 2 Xq L2 ( q ) (Lq (vq ) h) ; (Bq pq h) = (Kq + Fq h) h 2 Xq (Bq vq ) = 0 2 L2( q ):

(6.3.12) (6.3.13) (6.3.14) (6.3.15)

and consider the following problem: Find a pair of functions vq , pq satisfying

The problems (6.3.2){(6.3.5) and (6.3.13){(6.3.15) are equivalent provided that the equalities (6.3.2), (6.3.5) are considered in the distribution sense. By (6.3.1), (6.3.6), and Theorem 1.14.3, for each q 2 M, the function u ! u Pq is an isomorphism of X onto Xq and of L2( ) onto L2 ( q ). So, the problem (6.3.13){ (6.3.15) is equivalent to the following one: Find a pair of functions vq , pq such ~ ~ that (~q pq ) 2 X L2 ( ) v ~ (Lq (~q Pq ) h Pq ) ; (Bq (~q Pq ) h Pq ) = (Kq + Fq h Pq ) v p (Bq (~q Pq ) v

h2X

(6.3.16)

Pq ) = 0

2 L2 ( ):

(6.3.17) (6.3.18)

By Theorem 6.1.1 and Remark 6.1.1, there exists a unique solution of the problem (6.3.13){(6.3.15), so there exists a unique solution of the problem (6.3.16){ (6.3.18), and vq = vq Pq;1 , pq = pq Pq;1. ~ ~ Let us de ne functions Q1 and Q2 that map M into X by (Q1 (q) h) = (Kq h Pq ) We suppose (Q2 (q) h) = (Fq h Pq )

h 2 X:

(6.3.19) (6.3.20) (6.3.21)

Q1 and Q2 are continuous mappings of M into X , q ! Pq;1 is a continuous mapping of M into C 1 ( )n .

e e For q 2 M, we de ne operators Lq : X ! X and Bq 2 L(X L2 ( ) ) as follows: e (Lq (u) h) = (Lq (u Pq ) h Pq ) e (Bq u ) = (Bq (u Pq ) Pq )

u h2X u 2 X 2 L2 ( ) :

(6.3.22) (6.3.23)

6.3. Continuity with respect to the shape of the domain

453

e e 6.3.2 Lemmas on operators Lq and Bq

Lemma 6.3.1 Let a function ' satisfy the conditions (6.1.11){(6.1.14). Assume
e that (6.3.1), (6.3.6), (6.3.21) hold, and the operator Lq is de ned by (6.3.22). Then, for each q 2 M, there exist positive constants 1 (q), 2 (q) such that e e (Lq (u) ; Lq (w) u ; w) e e kLq (u) ; Lq (w)kX
2 (q )ku ; wkX 2 1 (q )ku ; wkX

u w2X u w2X

(6.3.24) (6.3.25)

and if qm ! q0 in M, then there are positive constants c1 and c2 such that 1 (qm ) c1 , 2 (qm ) c2 for any m.

Proof. Taking (6.3.22) into account and applying Lemma 6.1.4, we obtain e e (Lq (u) ; Lq (w) u ; w) 1 ku Pq ; w Pq k2 q X e e (6.3.26) j(Lq (u) ; Lq (w) h)j 2 ku Pq ; w Pq kXq kh Pq kXq u w2X
where the constants 1 , 2 depend only on the function '. Since the mapping u ! u Pq is an isomorphism of X onto Xq , we obtain (6.3.24), (6.3.25) from (6.3.26). The constants 1 (q), 2 (q) depend on the derivatives of the mapping Pq;1 . So, if qm ! q0 in M, then due to (6.3.21) there exist positive constants c1 and c2 such that 1 (qm ) c1 , 2 (qm ) c2 for each m, which completes the proof. Lemma 6.3.2 Let a function ' satisfy the conditions (6.1.11){(6.1.14). Assume
that (6.3.1), (6.3.6), (6.3.21) hold and

qm ! q0 vm ! v0
Then

in M in X: in X :

(6.3.27) (6.3.28) (6.3.29)

e e Lqm (vm ) ! Lq0 (v0 ) e the operator Lq :

Proof. By virtue of (6.3.10), (6.3.22), we obtain the following representation of

e (Lq (u) h) = 2 Z

'((Iq (u))(x))("qij (u))(x)("qij (h))(x) det((Pq;1 )0 (x)) dx

(6.3.30) (6.3.31)

where

@u @u ("qij (u))(x) = 1 @x i (x) @Pqk (Pq;1 (x)) + @xj (x) @Pqk (Pq;1 (x)) 2 @y @y
k j k i

454

Chapter 6. Optimization problems for steady ows

:::

(Iq (u))(x) =

n X i j =1

("qij (u))(x)

Pq = (Pq1 : : : Pqn )

(6.3.32)

(Pq;1 )0 (x) is the Frechet derivative of the mapping Pq;1 at point x 2 , and y = (y1 : : : yn ) = Pq;1 (x) 2 q . Denote the components of the mapping Pq;1 by Tqi , i.e.,

Pq;1 = (Tq1 : : : Tqn ): @Pqk (P ;1 (x)) = z (x)a (x) q qik @yi q

where zq (x) = det((Pq;1 )0 (x)) of the matrix (Pq;1 )0 (x). Denote
;
;1

(6.3.33)

The formula connecting derivatives of the components of a bijection with those of its inverse gives (see (1.14.8)) (6.3.34)

, aqik (x) is the cofactor of the element @Tqi (x) @xk (6.3.35) (6.3.36)

'm (x) = ' (Iqm (vm ))(x) "mij (x) = ("qm ij (vm ))(x) e ;1 0 Pm(x) = det((Pqm ) (x)) mki (x) = zqm (x)aqm ik (x) m = 0 1 2 ::: :
By (6.3.21), (6.3.27), (6.3.30){(6.3.36), we get
e e j(Lqm (vm ) ; Lq0 (v0 ) h)j = 2 ('m Pm "mij ; '0 P0 "0ij )"qm ij (h) dx e e Z

+2 where
m=

'0 P0 "0ij ("qm ij (h) ; "q0 ij (h)) dx e (c m + m kv0 kX )khkX

('m Pm "mij ; '0 P0 "0ij )2 dx e e
1 2

(6.3.37)

Z X n

i j =1

(6.3.38)

and lim m = 0: The triangle inequality yields

m
1m + 2m

(6.3.39) (6.3.40)

6.3. Continuity with respect to the shape of the domain

455

where
1m = 2m =

1 n 2 X (" ; " )2 dx 2 ('m Pm ) emij e0ij i j =1 1 Z n 2 2 X "2 dx : ('m Pm ; '0 P0 ) e0ij i j =1

(6.3.41) (6.3.42)

By (6.3.34), (6.3.36), and the triangle inequality,

@vmi (x) @Pqm k (P ;1 (x)) ; @v0i (x) @Pq0 k (P ;1 (x)) 2 dx 2 @xk @yj qm @xk @yj q0 (6.3.43) 1 Z 2 2 @vmi @v0i = 1m + 2m @xk mkj ; @xk 0kj dx where vmi , v0i are the components of the vector functions vm and v0 , and
1m = 2m =

Z Z

@vmi ; @v0i 2 2 dx 2 @xk @xk mkj 1 @v0i 2 ( 2 dx 2 : mkj ; 0kj ) @xk

0kj

(6.3.44)

The relations (6.3.21) and (6.3.27) give

mkj

in C ( ) as m ! 1:

(6.3.45)

So, by (6.3.28), we obtain 1m ! 0, 2m ! 0, and the right hand side of (6.3.43) also tends to zero. Now, taking into account (6.3.31), (6.3.41), and the estimate

k'm Pm kC (

8m

we obtain that 1m ! 0. By (6.3.21), (6.3.27), (6.3.28), and the Lebesgue theorem, we obtain 2m ! 0. At last, (6.3.29) follows from (6.3.37), (6.3.39), and (6.3.40).

Lemma 6.3.3 Let the conditions (6.3.1), (6.3.6), and (6.3.21) be satis ed and let e e an operator Bq be de ned by (6.3.11), (6.3.23). Then, the function q ! Bq is a continuous mapping of M into L(X L2 ( ) ). Proof. Due to (6.3.11), (6.3.23), we infer the following representation of the ope erator Bq : e (Bq u ) = Z

@ui (x) @Pqk (P ;1 (x)) (x) det((P ;1 )0 (x)) dx q @xk @yi q

(6.3.46)

456

Chapter 6. Optimization problems for steady ows

:::

where y = Pq;1 (x) 2 q . Suppose that in M: By using (6.3.34) and the notations (6.3.36), we get
e e ((Bqm ; Bq0 )u ) =

qm ! q0
Z

(6.3.47)

@ui @xk ( mki Pm ; 0ki P0 ) dx c m kukX k kL2( )

(6.3.48)

where

e e It follows from (6.3.47) and (6.3.21) that m ! 0, and (6.3.48) gives Bqm ! Bq0 in L(X L2 ( ) ). The lemma is proved.

m = k imax n k mki Pm ; 0ki P0 kC ( ) : =1 :::

By applying (6.3.19), (6.3.22), and (6.3.23), we can represent the problem (6.3.16){ (6.3.18) in the following form (~q pq ) 2 X L2 ( ) v ~ (6.3.49) e v e ~ (Lq (~q ) h) ; (Bq pq h) = (Q1 (q) + Q2 (q) h) h2X (6.3.50) e ~ (Bq vq ) = 0 2 L2 ( ): (6.3.51) Theorem 6.3.1 Let a function ' satisfy the conditions (6.1.11){(6.1.14). Assume that (6.3.1), (6.3.6), (6.3.20), (6.3.21) hold. Then, the function q ! (~q pq ) is a v ~ continuous mapping of M into X L2 ( ). Proof. 1) Suppose that qm ! q0 in M: (6.3.52) We take q = qm and h = vqm in (6.3.17). Then, from (6.1.12), (6.3.10), and ~ (6.3.18), we obtain 2a1kvqm Pqm kXqm kKqm + Fqm kXqm : ~ (6.3.53) This estimate, together with (6.3.20), (6.3.21), (6.3.52), gives kvqm kX c 8m: ~ (6.3.54) By (6.3.17), (6.3.52), and (6.3.53) kBqm (~qm Pqm )kXqm c1 8m: p (6.3.55)

6.3.3 Theorem on continuity

6.3. Continuity with respect to the shape of the domain

457

By Lemma 6.1.1

k(Bqm );1 kL(Vq0m L2 (

where

qm ))

(6.3.56) (6.3.57)
m:

Vq0m = f j f 2 Xqm (f u) = 0 u 2 Vqm

with Vqm de ned by (6.3.8), and also
m=
2L2 (

inf

) v2X

( P sup kv (Bqmkv Pqm ) P k qm ) Pqm Xqm k qm L2 ( qm )

This estimate and (6.3.21), (6.3.52) imply the existence of a positive constant ~ such that m ~. Thus, we can consider that m ~ for all m. Now, (6.3.55), (6.3.56), and (6.3.21) give

kpqm kL2 ( ~
vq ! v0 ~ ~ pq ! p0 ~ ~ e v Lq (~q ) !

c2

8m:

(6.3.58)

By (6.3.54), (6.3.58), and Lemma 6.3.1 (see (6.3.25)), we can extract from the sequence fvqm pqm Lqm (~qm )g a subsequence fvq pq Lq (~q )g such that ~ ~ e v ~ ~ e v weakly in X weakly in L2 ( ) weakly in X : in X (6.3.59) (6.3.60) (6.3.61) (6.3.62)

The relations (6.3.20) and (6.3.52) yield

Qi (q ) ! Qi (q0 )
By de nition, we have

i = 1 2: h2X

e v e ~ (Lqm (~qm ) h) ; (Bqm pqm h) = (Q1 (qm ) + Q2 (qm ) h) e ~ (Bqm vqm ) = 0 2 L2 ( ):

(6.3.63) (6.3.64) (6.3.65)

Applying (6.3.52), (6.3.59), (6.3.64), and Lemma 6.3.3 gives

e ~ lim (Bq vq !1 e ~ ) = (Bq0 v0 ) = 0

2 L 2 ( ):

By (6.3.60){(6.3.62) and Lemma 6.3.3, we pass to the limit in (6.3.63), with m changed by . This gives
e ~ ; Bq0 p0 = Q1 (q0 ) + Q2 (q0 ):

(6.3.66)

458

Chapter 6. Optimization problems for steady ows

:::

Lemma 6.3.1 yields

e v e (Lq (~q ) ; Lq (u) vq ; u) 0 ~

u2X 8 :

(6.3.67)

e ~ ~ Taking into account that (Bq pq vq ) = 0, by (6.3.59), (6.3.62), and (6.3.63), we get e v (Lq (~q ) vq ) = (Q1 (q ) + Q2 (q ) vq ) ! (Q1 (q0 ) + Q2 (q0 ) v0 ): ~ ~ ~ The relations (6.3.61) and (6.3.66) give e v e ~ lim (Lq (~q ) u) ; (Bq0 p0 u) = (Q1 (q0 ) + Q2 (q0 ) u) !1

(6.3.68) (6.3.69) (6.3.70)

u 2 X:

It follows from (6.3.52) and Lemma 6.3.2 that

e e Lq (u) ! Lq0 (u)

in X

u 2 X: u 2 X:

By (6.3.59), (6.3.68){(6.3.70), we pass to the limit in (6.3.67). By virtue of (6.3.65), we get Take here u = v0 ; h, > 0, h 2 X , and let tend to zero. Then, because of the ~ e continuity of the mapping Lq0 (see (6.3.25)), we obtain
e e ~ ~ (Q1 (q0 ) + Q2 (q0 ) ; Lq0 (u) + Bq0 p0 v0 ; u) 0

(6.3.71)

h 2 X: It follows from the latter estimate and (6.3.65) that the pair vq0 = v0 , pq0 = p0 ~ ~ ~ ~ is a solution of the problem (6.3.49){(6.3.51) for q = q0 .
2. By using (6.3.59), (6.3.68){(6.3.70), is easy to see that
e v e v ~ (Lq (~q ) ; Lq (~q0 ) vq ; vq0 ) ! 0: ~ Thus, Lemma 6.3.1 gives vq ! vq0 strongly in X . Since the solution of the problem ~ ~ (6.3.16){(6.3.18) is unique, we obtain

e v e ~ (Q1 (q0 ) + Q2 (q0 ) ; Lq0 (~0 ) + Bq0 p0 h) 0

vqm ! vq0 ~ ~

strongly in X:

(6.3.72)

Subtracting the equality (6.3.63) from the one for m = 0 gives

e ~ e ~ Bqm pqm ; Bq0 pq0 = m in X e v e v m = Lqm (~qm ) ; Lq0 (~q0 ) ; Q1 (qm ) ; Q2 (qm ) + Q1 (q0 ) + Q2 (q0 ):

(6.3.73)

By virtue of (6.3.20), (6.3.47), (6.3.72), and Lemma 6.3.2, we get

m!0

in X :

(6.3.74)

6.4. Control of uid ows by perforated walls

:::

459

Further,

Lemma 6.3.3, (6.3.52), and (6.3.58) yield m ! 0. So, (6.3.73){(6.3.75) give e p Bq0 (~qm ; pq0 ) ! 0 in X : ~ Finally, (6.3.23) and Lemma 6.1.1 imply pqm ! pq0 in L2 ( ). ~ ~

e ~ e ~ e e p e p Bqm pqm ; Bq0 pq0 = (Bqm ; Bq0 )~qm + Bq0 (~qm ; pq0 ) ~ e e p e e k(Bqm ; Bq0 )~qm kX kBqm ; Bq0 kL(L2( ) X ) kpqm kL2( ) = m : ~

(6.3.75)

6.4 Control of uid ows by perforated walls and computation of the function of ltration
The solution of the problem (6.2.2){(6.2.4) on the ow of a nonlinear viscous, in particular, viscous uid, depends on the functions of volume and surface forces K , F , and on the functions of viscosity ' and ltration . The viscosity function may be varied by the introduction of various additions in the uid. In melted metals, the volume forces may be created by electromagnetic elds. For such problems, one should solve a coupled system of equations for the uid and the electromagnetic eld. The electrical resistance of ordinary uids is high, and so they interact very faintly with electromagnetic elds. However, created were many suspensions which interact with them. Thus, volume forces in these suspensions may be originated by electromagnetic elds. For these problems, we have to solve a coupled system of equations for the uid and the electromagnetic eld. So, in what follows, we do not consider the volume forces and the function of viscosity as controls. Nevertheless, they may be enclosed in the optimization problems under consideration. In order to control the ow of the uid, one uses perforated walls, which may be placed at either the inlet, or outlet, or inside of the canal. In Fig. 6.4.1, a spraying device (a header) of a paper machine is schematically shown. S Here, we have S = 4=1 S i , where S is the boundary of the domain de ned i by the shape of the header, S1 the in ow, S2 the hard wall, S3 , S4 are the out ows. The canal has a perforated wall at S3 , and there is no wall at S1 and S4 . The space of S4 is small, and so there is a small out ow here, which is used in practice as a control parameter. We will now show that the perforated wall may be modeled by a ltration layer.

6.4.1 The problem on ow in a circular cylinder and the function of ltration

We suppose that the holes of the perforated wall are cylinders of a circular crosssection. Let us consider the problem on ow of a nonlinear viscous uid in a

460

Chapter 6. Optimization problems for steady ows

S1

:::

S2

W
S2 S3 S4

Figure 6.4.1: A scheme of a header of a paper machine cylinder of the wall. We apply cylindrical coordinate system (r z ) and suppose that the velocity vector v = (vr v vz ) is such that v = 0 and vr , vz are functions of r, z only, see Fig. 6.4.2. Then, the following components of the stress tensor are not equal to zero:
rr = ;p + 2'(I (v ))

@vr @r vr = ;p + 2'(I (v)) r

2

zz = ;p + 2'(I (v ))

@vz @z @vz + @vr rz = '(I (v )) @r @z

2

(6.4.1) (6.4.2) (6.4.3) (6.4.4) (6.4.5)

where

2 1 + 2 @vz + @vr : @r @z The equations of motion and incompressibility are the following:

I (v) = @vr @r

+ vrr

+ @vz @z

Here, K = (Kr Kz ) is the function of volume forces. In many cases, one can put K = 0. As the radii of the cylinders of a perforated wall are of orders smaller than the characteristic dimensions of the wall, we can consider that the velocity pro le at the in ow of the cylinder is close to a rectangular, i.e., vr = 0 vz = bf (r) at z = 0 (6.4.6) where b = const > 0 and f is a function continuous on 0 R], f (r) = 1 for an arbitrary r 2 0 R ; ], f (R) = 0, R being the radius of the cylinder, a small positive constant.

@ rr + @ rz + rr ; = Kr @r @z r @ rz + @ zz + rz = K z @r @z r @vr + @vz + vr = 0: @r @z r

6.4. Control of uid ows by perforated walls

v

:::

461

vr
v
vz

Figure 6.4.2: Notation for ow in a circular cylinder

b b (Fr Fz ): b At the out ow of the cylinder, we prescribe the following surface forces F = b Fr = rz z=l = 0 b Fz = zz z=l = = const 0

(6.4.7) l being the length of the cylinder. In addition, the condition of adhesion gives v = 0 at r = R: (6.4.8) We denote by Q the mean normal force at the in ow of the cylinder, and by v the mean velocity in the tube: ^ ZR Q=; 2 (r 0)r dr (6.4.9)
zz 0 ZR v = R2 vz (r z )r dr: ^ 2 0

R2

(6.4.10)

By virtue of the condition of incompressibility (6.4.5), the latter integral is independent of z , and so, by (6.4.6), we have ZR f (r)r dr: (6.4.11) v = 2b ^ Suppose the problem (6.4.1){(6.4.8) is solved, its solution may be computed numerically. Then, we can compute Q and v by (6.4.9), (6.4.10). Let us de ne a ^ passage factor of the tube, , by ^ = Qv : (6.4.12) +

R2

462

Chapter 6. Optimization problems for steady ows

:::

Obviously, depends on v , R, l, and it is almost independent of . Note that, ^ for the case of a rectilinear ow, is independent of , see the formulas (6.4.22), (6.4.23) below and Litvinov (1982a). Thus, we denote by (^ R l). v Let now G be a perforated wall which is considered as an (n ; 1)-dimensional manifold, particularly, G = S3 for the header shown in Fig. 6.4.1. Let M be the number of holes (cylinders) in the wall. Assume they are numbered and Gi G is the region of the ith hole, i = 1 : : : M , and Ri , li are the radius and length of the ith cylinder. We de ne a passage function ~ of the wall G as follows:

R+ G 3 ( s) ! ~( s) = 0 (

1 2

Ri li ) if s 2 Gi S= 1 : : : M (6.4.13) i if s 2 G n M Gi : i=1

Comparing (6.4.12) and (6.4.13) with the boundary condition (6.1.9), we can introduce a ltration function ~ given on R+ G by 8 1 1 = < ( 1 R l ) if s 2 Gi i = 1 : : : M ~( s) = 2 i i (6.4.14) S ~( s) :1 if s 2 G n M Gi : i=1 The number M is usually very large in practice, and the radii of the cylinders are of orders smaller than the characteristic dimensions of the perforated wall. Therefore, ~ is a rapidly oscillating function which can take value 1. Such functions are inconvenient for computation. This is why we will now introduce an averaging ltration function. An averaged passage function is de ned by ( s) =
Z

! (s ; )~ (

)d

(6.4.15)

where ! is an averaging kernel (see Section 1.6.4). The formula (6.4.15) determines the function everywhere in G except for a slender band of width , and it may be extended to this band. Now, de ne an averaging ltration function by ( s) = ( 1 s ) : (6.4.16)

6.4.2 The passage factor for the power model

The viscosity function for the power model is given by 2 2 R+ '( ) = a1 k;1 (6.4.17) where a1 , k are positive constants. If k = 1, this is just the case of a viscous, Newtonian uid. For the power model, the problem on ow in a circular tube under the conditions that vr = v = 0, where v is the tangential component of

6.4. Control of uid ows by perforated walls

:::

463

where p0 = p(0), p1 = p(l). Now, zz = ;p and (6.4.7) and (6.4.9) give = ;p1 , Q = p0 . Thus, (6.4.19) yields

v, is exactly solved, see, e.g., Astarita and Marrucci (1974), in this case, vz is a function of r only, and p is an a ne function of z , i.e., 1 k k k vz (r) = k + 1 21a dp k R k+1 ; r k+1 (6.4.18) dz 1;k (6.4.19) a = 2 2 a1 p = p0 + (p1 ; p0 ) z l dp = ; Q + : dz l

(6.4.20)

By substituting (6.4.18) into (6.4.10), we get

dp 1 k v = 3k k 1 21a dz k R k+1 : ^ +
From (6.4.20) and (6.4.21) + k^ Q + = R2kal 3k k 1 vk : +1
k+1 k ^ (^ R l) = Q v = R al 3k k 1 v1;k : v + 2 + ^

(6.4.21)

(6.4.22)

By (6.4.12), (6.4.22), we obtain the following formula for the passage factor of the power model (6.4.23)

By using (6.4.13), (6.4.15), (6.4.16), and (6.4.23), one can compute the ltration function for the power model.

In some cases, it is necessary to obtain a required distribution of the surface forces. We will now show that, under some restrictions, the required distribution of the normal component of the surface forces at the inlet may be obtained by the ltration layer (perforated wall). Let a uid ow in the canal shown in Fig. 6.4.3, where is the domain of S the canal, S is the boundary of , S = 3=1 S i , Si 's are open sets in S , S1 is the i hard wall, S2 the inlet, and S3 the outlet. The ltration layer is attached to the e e e inlet. We suppose that the function of surface forces F = (F1 : : : Fn ) acting on 0 the external surface of the ltration layer S2 is known. The functions of velocity v and pressure p satisfy the equations (6.1.6). Suppose we are given also a function

6.4.3 Control of the surface forces at the inlet by the perforated wall

464 wall, i.e.,

Chapter 6. Optimization problems for steady ows

:::

P = (P1 : : : Pn ) of the surface forces at the outlet and zero velocities on the hard v S1 = 0 ; p ij + 2'(I (v))"ij (v) j S3 = Pi i = 1 ::: n
(6.4.24) (6.4.25)

i being the components of the unit outward normal to S . Notice that, instead of (6.4.25), the conditions (6.1.9), (6.1.10) may be considered. In the latter case, ^ ^ ^ the functions F = (F1 : : : Fn ) and are supposed to be known.
S1

~ F

W
S2 S2

S3 S1

Figure 6.4.3: Domain at the inlet

of the canal with ltration layer (perforated wall) placed

On S2 , we have the following conditions of ltration (see (6.1.9), (6.1.10))

; p ij + 2'(I (v))"ij (v) j (vi ; v i )

e i S2 ;Fi i S2 = ; ~ (v 2 )v i = 1 : : : n: S2 = 0

S2

(6.4.26) (6.4.27)

Here, we suppose that the thickness of the ltration layer is small, and carry the 0 e e function F from S2 to S2 , see Fig. 6.4.3. We also assume that F 2 C (S 2 )n . ~ as an unknown function that should be calculated so that the Consider normal component of the function of surface forces is equal to q on S2 , i.e.,

; p ij + 2'(I (v))"ij (v)

e q(s) ; Fi (s) i (s)

j i S2 = q

(6.4.28) (6.4.29)

where q is a continuous function on S2 . In this case, we suppose that

s 2 S2

a small positive constant. Consider the problem: Find a pair of functions v, p satisfying the equations (6.1.6) and the boundary conditions (6.4.24), (6.4.25), (6.4.27), (6.4.28). By analogy with the proved above, see Section 6.1, it may be shown that, if K 2 L2 ( )n , P 2 L2 (S3 )n , q 2 C (S 2 ), then there exists a unique solution of this problem such that v 2 H 1 ( )n , p 2 L2 ( ). We denote it by v(q), p(q). Let v(q) be the normal

6.5. The ow in a canal with a perforated wall placed inside

465

component of v(q), i.e., v(q) = v(q)i i . Suppose that v(q) < 0 a.e. on S2 , the in; let. Now, we de ne a ltration function ~ : R+ S2 ! R at points (v(q) (s))2 s , s 2 S2 , by
e ~((v(q) )(s))2 s) = ; q(s) ; Fi (s) i (s) v(q) (s)

s 2 S2

(6.4.30)

It follows from (6.4.28) and (6.4.30) that the pair v(q), p(q) satis es (6.4.26). Thus, if the ltration layer meets (6.4.30), then (6.4.28) is satis ed.

6.5 The ow in a canal with a perforated wall placed inside

In Section 6.1, we considered the problem on ow of a nonlinear viscous uid in the case when the ltration layer (perforated wall) was put on the boundary of the canal. Now, we will consider the case of the ltration layer placed inside of the canal. Let be a bounded domain with a Lipschitz continuous boundary S . Suppose and 2 are open subsets of such that = 1 2 , 1 \ 2 = ?, and 1 ; = 1 \ 2 is an (n;1)-dimensional, Lipschitz continuous manifold., see Fig. 6.5.1.
S1
V

6.5.1 Basic equations

S2

S3

^ F

S1

Figure 6.5.1: Notation for ow in a canal with ltration layer ; placed inside of it We assume that S1 , S2 , S3 are open subsets of S such that S = 3=1 S i , i Si \ Sj = ? for i 6= j . We consider ; and S3 as ltration layers and S1 as a hard wall. Let us consider the following problem: Find a pair of functions v, p (velocity and pressure) satisfying the equations of motion and incompressibility:
S

@p @ ; @xi ; 2 @xj '(I (v))"ij (v) = Ki div v = 0

in in

1 1

2 2

i = 1 ::: n

(6.5.1) (6.5.2)

466

Chapter 6. Optimization problems for steady ows

:::

and the following boundary conditions on Si :

v S1 = 0 ; p ij + 2'(I (v))"ij (v) j S2 = Fi i = 1 ::: n ^ ; p ij + 2'(I (v))"ij (v) j i S3 ;Fi i S3 = ; (v2 )v S3 (vi ; v i ) S3 = 0 i = 1 ::: n

(6.5.3) (6.5.4) (6.5.5) (6.5.6)

where = ( 1 : : : n ) is the unit outward normal to S , v = vk k . ( ( Put v(i) = v, p(i) = p in i , (i) = ( 1i) : : : ni) ) is the unit normal to ; ( ( directed outwards i , v(i) = vki) ki) on ;. The boundary conditions of ltration on ; are the following:

+ ; p(2)

v(1) ; = v(2) ; (6.5.7) (1) ; v (1) (1) ) = 0 (v (6.5.8) ; (1) + 2'(I (v (1) ))" (v (1) ) (1) (1) ; p ij ij j i ; (2) ))" (v (2) ) (2) (1) = ; ~ ((v (1) )2 )v (1) (1) : (6.5.9) ij + 2'(I (v ij j i ; i i
(v(2) ; v(2)
(2) ) ; = 0:

It follows from (6.5.7) and (6.5.8) that

6.5.2 Generalized solution of the problem

De ne spaces X and V as follows:

X = u j u 2 H 1 ( )n u S1 = 0 (u ; u ) S3 = 0 (u ; u(1) V = u j u 2 X div u = 0 :

(1) )

;= 0

(6.5.10) (6.5.11)

Note that, in (6.5.10), the last equality is equivalent to (u ; u(2) (2) ) ; = 0. The expression (6.1.21) de nes norms in X , V that are equivalent to the norm of H 1 ( )n . X and V are Hilbert spaces with the scalar product (u h) = Introduce a set B3 as follows:
Z

"ij (u)"ij (h) dx:

B3 =

j 2 L1(; C 1 (R+ ))

satis es the conditions (6.1.16){(6.1.18) with positive constants b1 , b2 for an arbitrary ( s) 2 R+ ; : (6.5.12)

6.5. The ow in a canal with a perforated wall placed inside

467

This set is equipped with the topology generated by that of L1 (; C 1 (R+ )). Suppose that ' 2 B1 , 2 B2 , ~ 2 B3 , B1 and B2 de ned by (6.2.5), (6.2.8). De ne also operators L : X ! X and B 2 L(X L2 ( )) by (L(u) h) = 2 +
Z Z
;

'(I (u))"ij (u)"ij (h) dx

~(juj2 )ui hi ds +
Z

P Here, juj2 = n=1 u2 . We suppose i i

Bu = div u:

S3

(juj2 )ui hi ds

(6.5.13)

K = (K1 : : : Kn ) 2 X F = (F 1 : : : F n ) 2 X ^ ^ ^ F = (F1 : : : Fn ) 2 X :
^ In particular, if K 2 L2 ( )n , F 2 L2(S2 )n , F 2 L2(S3 )n , then (K h) =
Z

(6.5.14)
Z

Ki hi ds

(F h) =

h 2 X:

S2

Fi hi ds

^ (F h) =

S3

^ Fi hi ds (6.5.15)

The generalized solution of the problem (6.5.1){(6.5.9) is de ned as a pair of functions v, p satisfying (v p ) 2 X L 2 ( ) ^ (L(v) h) ; (B p h) = (K + F + F h) (Bv q) = 0 q 2 L2 ( )

h2X

(6.5.16) (6.5.17) (6.5.18)

By using Green's formula, one can show that, if v, p is a solution to the problem (6.5.16){(6.5.18), then v, p is a solution to the problem (6.5.1){(6.5.9) in the distribution sense. On the contrary, if v, p is a solution to (6.5.1){(6.5.9) that satis es (6.5.16), then v, p is a solution to (6.5.16){(6.5.18). By analogy with the above, see Section 6.1, one proves the following

Theorem 6.5.1 Suppose sets B1, B2, B3 are de ned by (6.2.5), (6.2.8), (6.5.12), and ' 2 B1 , 2 B2 , ~ 2 B3 . Assume also that operators L and B are de ned by

^ (6.5.13) and the functions of volume and surface forces K , F , F meet (6.5.14). Then, there exists a unique solution v, p of the problem (6.5.16){(6.5.18). Let fXk g, fMk g be sequences of nite-dimensional subspaces of X and L2 ( ) such that (6.1.37), (6.1.38), (6.1.42) hold, and Xk Xk+1 , Mk Mk+1 for each k. Then, for any k, there exists a unique solution vk , pk of the problem (6.1.57){ (6.1.59) and vk ! v in X , pk ! p in L2 ( ):

468

Chapter 6. Optimization problems for steady ows

:::

The solution of the problem (6.5.16){(6.5.18) depends on the functions of ^ loading K , F , F and on the functions of viscosity and ltration ', , ~, so that we ^ ^ denote the solution by v(K F F ' ~), p(K F F ' ~). By Theorem 6.5.1, the following function G is well de ned: ^ ^ (X )3 B1 B2 B3 3 (K F F ' ~) ! G(K F F ' ~) ^ ^ = (v(K F F ' ~) p(K F F ' ~)) 2 V L2( ): (6.5.19) By analogy with the proof of Theorem 6.2.1, we get Theorem 6.5.2 Suppose a function G is given by (6.5.19) and sets B1, B2, B3 are de ned by (6.2.5), (6.2.8), (6.5.12) and equipped with the topologies generated by those of C 1 (R+ ), L1(S3 C 1 (R+ )), and L1(; C 1 (R+ )), respectively. Then, the function G is a continuous mapping of (X )3 B1 B2 B3 into V L2( ):

6.6 Optimization by the functions of surface forces and ltration

We again consider the case of ow studied in the previous section, i.e., the problem (6.5.16){(6.5.18) with L de ned by (6.5.13). We assume the functions of surface ^ forces F , F and the functions of ltration , ~ are controls, while the functions of viscosity ' and volume forces K are xed, and

6.6.1 Formulation of the problem and the existence theorem

' 2 B1

K2X

(6.6.1)

B1 de ned by (6.2.5). As shown in Section 6.4, the functions of ltration and surface

forces may be varied by changing the number of the holes in the perforated wall, their dimensions and dispositions. Thus, the number of the holes, the coordinates of the centers of the holes, and their diameters can be considered as controls. But it is much more convenient to consider , ~ as controls which are functions of points of S3 and ; only. Such an approach does not diminish the generality. ^ For if 0 , ~0 are the optimal ltration functions given on S3 and ;, F0 , F0 are the optimal functions of surface forces, and v0 , p0 is the solution of the problem ^ ^ (6.5.16){(6.5.18) for = 0 , ~ = ~0 , F = F0 , F = F0 , then we can put ~1 (jv0 So,
1 1 (jv0 (s)j 2

(s)j2

s) = 0 (s) s) = ~0 (s)

s 2 S3 s 2 ;:

(6.6.2) (6.6.3)

and ~1 are the optimal functions de ned at points (jv0 (s)j2 s), s 2 S3 and s 2 ;. The perforated walls may be computed so as to obtain good approximations of the functions 1 and ~1 , see Section 6.4.

6.6. Optimization by the functions of surface forces and ltration

469

We introduce a set of controls by 1 1 ^ N = T j T = (F F ~) 2 H ; 2 (S2 )n H ; 2 (S3 )n C (S 3 ) C (;) (6.6.4) (s) 0 s 2 S 3 ~(s) 0 s 2 ; : The set N is equipped with the topology generated by that of the product 1 2 H ; 1 (S2 )n H ; 2 (S3 )n C (S 3 ) C (;): Suppose we are given functionals i satisfying the condition (6.6.5) i are functionals continuous on N X L2 ( ), i = 0 1 : : : m. By Theorem 6.5.1, for each T 2 N , there exists a unique solution of the problem (6.5.16){(6.5.18), which will be denoted by v(T ), p(T ). On the set N , we de ne functionals i as follows: i = 0 1 : : : m: (6.6.6) i (T ) = i (T v (T ) p(T )) Determine a set of admissible controls by ^ Nad = T j T = (F F ~) 2 N F = (F1 : : : Fn ) 2 L2(S2 )n kF kL2(S2 )n c1 ^ ^ ^ ^ F = (F1 : : : Fn ) 2 L2 (S3 )n kF kL2 (S3 )n c2 2 Wpl (S3 ) l (;) k ~ k l k kWpl (S3 ) c3 ~ 2 Wp Wp (;) c4 p > 1 lp > n ; 1 i (T ) 0 i = 1 : : : r i (T ) = 0 i = r + 1 : : : m r m : (6.6.7) Moreover, we suppose that S3 and ; are (n ; 1)-dimensional manifolds of the C l class. The optimization problem consists in nding T0 such that ^ T0 = (F0 F0 0 ~0 ) 2 Nad inf (6.6.8) 0 (T0 ) = T 2N 0 (T ):

Theorem 6.6.1 Suppose the conditions (6.6.1), (6.6.5), and (6.6.6) are satis ed

ad

and a nonempty set Nad is de ned by (6.6.7). Then, there exists a solution of the problem (6.6.8). Proof. It follows from Theorem 6.5.2 that the function N 3 T ! (v(T ) p(T )) 2 V L2 ( ) is continuous. Hence, i , i = 0 1 : : : m, are functionals continuous on N , and Nad is a compactum in N . So, there exists a solution of the problem (6.6.8). Let us consider some forms of the functionals i . The goal functional can be taken in the form
0 (T ) =

Z X n

S3 i=1

(v(T )i ; zi )2 ds

(6.6.9)

470

Chapter 6. Optimization problems for steady ows

:::

where zi 's are given functions from L2 (S3 ). Hence, in this case, the velocity eld at the out ow should be close to z = (z1 : : : zn). Particularly, for the header of the paper machine and for the extrusion head, the function z may be taken in the form

zi i = const > 0 zk ; zi i k = 0 k = 1 : : : n: Thus, z is normal to S3 and z is a constant vector.

0 (T ) =

(6.6.10)

In some cases, it is necessary to obtain a velocity eld that is close to some given one in the domain . The goal functional is taken then in the form
Z

n X i=1

(v(T )i ; yi )2 dx

(6.6.11)

where yi 's are prescribed functions from L2( ). The functional of restriction 1 may be de ned by
1 (T ) =

Z X n

i j =1

'(I (v(T ))) "ij (v(T )) 2 dx ; c:

(6.6.12)

The condition 1 (T ) 0 means now that the power must not exceed c. Next, we are going to consider the optimization problem (6.6.8) in which the state of the system v(T ), p(T ) is calculated approximately by the Galerkin method. To this end, we will derive the Frechet derivatives of the functionals i and necessary optimality conditions. So, let us deduce now the Frechet derivatives of the function T ! (v(T ) p(T )). De ne an operator L1 : X ! X by (L1 (u) h) = 2
Z

6.6.2 On the di erentiability of the function T ! (v(T ) p(T ))

'(I (u))"ij (u)"ij (h) dx u h 2 X:
(6.6.13)

Let fXk g be a sequence of nite-dimensional subspaces of X that satis es (6.1.37) and

1 Xk 2 H1 ( )n

8k:

(6.6.14)

Lemma 6.6.1 Let a function ' satisfy the conditions (6.1.11){(6.1.14). Let also fXk g be a sequence of nite-dimensional subspaces of X satisfying (6.6.14). Then,

for any k, the operator L1 considered as a mapping of Xk into Xk is continuously Frechet di erentiable in Xk , and at any point u 2 Xk its derivative L01 (u) is given by

6.6. Optimization by the functions of surface forces and ltration

471

(L01 (u)w h) = 2

Z h

'(I (u))"ij (w)"ij (h) i + 2 d' (I (u))"lm (u)"lm (w)"ij (u)"ij (h) dx d w h 2 Xk : (6.6.15)

Moreover,

(L01 (u)h h)

khk2 X

= 2 min(a1 a3 )

(6.6.16)

where a1 , a3 are positive constants from (6.1.12), (6.1.13).

1 Proof. In the subspace Xk , the norm of X is equivalent to that of H1( )n ,

because Xk is nite-dimensional and Xk conclude

1 H1 ( )n . So, by using (6.1.11), we

k'(I (u + w))"ij (u + w) ; '(I (u))"ij (u) ; '(I (u))"ij (w) ; 2 d' (I (u))"lm (u)"lm (w)"ij (u)kL1( ) !(w)kwkX d

u w 2 Xk

(6.6.17)

where !(w) ! 0 as kwkX ! 0. Thus, (6.6.15) follows from (6.6.17). By applying the estimate ; "ij (u)"ij (h) 2 I (u)I (h) we get (6.6.16) from (6.6.15) and (6.1.12){(6.1.14), see (6.1.54). Remark 6.6.1 One can show that the operator L1 considered as a mapping of X into X is G^teaux di erentiable in X and its G^teaux derivative is given by a a (6.6.15). The estimate (6.6.16) holds also for arbitrary u h 2 X . De ne a space U and a set N1 by
2 2 (6.6.18) U = H ; 1 (S2 )n H ; 1 (S3 )n C (S 3 ) C (;) ^ N1 = T j T = (F F ~) 2 U (s) > ; s 2 S 3 ~(s) > ; s 2 ; :

(6.6.19)

Here, is a small positive constant which is de ned so that (L01 (u)h h) +

Z
; i=1

~ X h2 ds + i

u h2X

2 C (S 3 ) > ; ~ 2 C (;) ~ > ;

n X 2 hi ds S3 i=1

2 1 khkX 1 = const > 0:

(6.6.20)
1

By virtue of (6.6.16) and the embedding theorem, there exists a constant which (6.6.20) holds.

for

472

Chapter 6. Optimization problems for steady ows

:::

We now determine a mapping Q : N1 X L2 ( ) ! X ^ T = (F F ~) 2 N1

; ( R R (L1 (u) h) + ; ~ui hi ds + S3 = R

L2 ( ) by

(u h) 2 X 2 Q(T u p) (h q) =

(p q) 2 L2 ( )2

q div u dx:

ui hi ds ;

^ p div h dx ; (K + F + F h) (6.6.21)

Lemma 6.6.2 Let (6.6.1) hold, let fXk g, fMk g be sequences of nite-dimensional
Q0 (T u p)(MT Mu Mp) = @Q (T u p)MT + @ (@Qp) (T u p)(Mu Mp) @T u
M

1 subspaces of X and L2 ( ), respectively, let Xk H1 ( )n for each k, and let (6.1.42) hold. Then, the operator Q considered as a mapping of N1 Xk Mk into Xk Mk is continuously Frechet di erentiable and its derivative is given by

@Q (T u p)MT (h q) = @T 0 @Q @ (u p) (T u p)(Mu Mp) (h q) ( R ~ R R 0 R = (L1 (u)Mu h) + ; Mui hi ds + S3 Mui hi ds ; q div Mu dx h 2 Xk q 2 Mk :

^ T = (MF MF

(R

u 2 Xk Mp 2 Mk ~ ^ ; M ui hi ds + S3 M ui hi ds ; (MF + MF h)
M M

~) 2 U

p div h dx
(6.6.22)

At any point (T u p) 2 N1 Xk Mk , the operator @ (@Qp) (T u p) is an isomoru phism of Xk Mk onto Xk Mk :

Proof. For an arbitrary xed (u p) 2 Xk Mk , the function T ! Q(T u p) is a linear continuous mapping of an open subset N1 of U into Xk Mk . So, the continuous Frechet di erentiability of the mapping Q : N1 Xk Mk ! Xk Mk
(w ) 2 Xk Mk

and the formulas (6.6.22) follow from (6.6.21) and Lemma 6.6.1. Let (y z ) be an arbitrary point of Xk Mk . Consider the problem: Find a pair of functions w, satisfying

@Q @ (u p) (T u p)(w ) = (y z )
where (T u p) 2 N1 Xk Mk .

(6.6.23)

6.6. Optimization by the functions of surface forces and ltration

473

Introduce operators L2 2 L(Xk Xk ) and Bk 2 L(Xk Mk ) by (L2 v h) = (L01 (u)v h) + (Bk v ) =

Z Z;

~vi hi ds + div v dx:

S3

vi hi ds

(6.6.24)

By (6.6.22), the problem (6.6.23) may be rewritten as follows: (w ) 2 Xk Mk (L2 w h) ; (Bk h) = (y h) h 2 Xk (Bk w q) = (z q) q 2 Mk : (6.6.25) (6.6.26) (6.6.27) (6.6.28) (6.6.29) (6.6.30) (6.6.31)

From Lemma 6.1.3, we infer the existence of a unique w0 2 Vk? such that

Bk w0 = z

kw0 kX

;1

kz kMk :
h 2 Vk :

Thus, the problem (6.6.25){(6.6.27) reduces to the following one: Find w1 satisfying

w1 = w ; w0 2 Vk (L2 w1 h) = (y h) ; (L2 w0 h)
From (6.6.20) and (6.6.24) (L2 h h)
2 1 khkX

h 2 Xk :

Thus, by the Lax{Milgram theorem (Theorem 1.5.2) and (6.6.28), we conclude the existence of a unique solution w1 of the problem (6.6.29), (6.6.30) and

kw1 kX c(kykXk + kz kMk )

(6.6.32)

where the constant c is independent of y, z , and k. Now, L2 w ; y 2 Vk0 (see (6.1.41)), therefore, according to Lemma 6.1.3, there exists a unique 2 Mk such that Bk = L2w ; y (6.6.33) ;1 kL w ; y k k kL2( ) 2 Xk c1 (ky kXk + kz kMk ): Thus, the operator @ (@Qp) (T u p) is an isomorphism of Xk Mk onto Xk Mk . u

Theorem 6.6.2 Let fXk g, fMk g be sequences of nite-dimensional subspaces of 1 X and L2( ), X de ned by (6.5.10). Let also Xk H1 ( )n for each k and (6.1.42), (6.6.1) hold. For an arbitrary xed k, de ne a function F : N1 ! (Xk Mk )
as follows :

474

Chapter 6. Optimization problems for steady ows

;

:::

^ T = (F F ~) 2 N1 F (T ) = v(T ) p(T ) 2 Xk Mk where v(T ), p(T ) is the solution of the problem (L1 (v(T )) h) +
Z

(6.6.34)

^ p(T ) div h dx = (K + F + F h)
Z

~v(T )i hi ds +

S3

v(T )i hi ds h 2 Xk q 2 Mk : @Q (T v(T ) p(T )) @T

(6.6.35) (6.6.36)

q div v(T ) dx = 0
;1

Then, F is a continuously Frechet di erentiable mapping of N1 into Xk Mk , and its derivative is given by

F 0 (T ) = ; @ (@Qp) (T v(T ) p(T )) u

;1

(6.6.37)

where @ (@Qp) (T v(T ) p(T )) is the inverse mapping of @ (@Qp) (T v(T ) p(T )): u u Proof. It follows from (6.6.21) and (6.6.34){(6.6.36) that F is the implicit function de ned by the equation Q(T F (T )) = 0 (6.6.38) where Q is considered as a mapping of N1 Xk Mk into Xk Mk . So, the theorem follows from Lemma 6.6.2 and the theorem on the di erentiability of an implicit function (Theorem 1.9.2). Remark 6.6.2 In Theorem 6.6.2, by using the implicit function theorem, we have proved the Frechet di erentiability of the function F : T ! F (T ) = (v(T ) p(T )), where v(T ), p(T ) is the Galerkin approximation of the solution of the problem (6.5.16){(6.5.18), i.e., the solution of the problem (6.6.34){(6.6.36) for an arbitrary xed k. The questions of both di erentiability of the function T ! (v(T ) p(T )), where v(T ), p(T ) is the solution of (6.5.16){(6.5.18), and necessary optimality conditions for this case are open yet. Apparently, it could happen to be useful to apply the concept of an extended di erentiability that is introduced and applied to the solution of optimization problems for nonlinear partial di erential equations in works by Serovaiskii (1991), (1993a, b).

6.6.3 Di erentiability of the functionals optimality conditions

and necessary

We suppose now that the functionals i satisfy the following condition, which is stronger than (6.6.5), i are functionals continuously Frechet di erentiable in (6.6.39) N1 X L2 ( ), i = 0 1 : : : m.

6.6. Optimization by the functions of surface forces and ltration

475

(6.6.34){(6.6.36), and (6.6.39) hold. Let functionals i be de ned by (6.6.6). Then, ^ i are continuously Frechet di erentiable in N1 and at any point T = (F F ~ ) 2 N1 the Frechet derivative of i is given by
i (T )MT =
0

Theorem 6.6.3 Let fXk g, fMk g be sequences of nite-dimensional subspaces of 1 X and L2( ), X de ned by (6.5.10). Let also Xk H1 ( )n for each k and let (6.1.42), (6.6.1) hold. Suppose that v(t), p(T ) is the solution to the problem
@ i (T v(T ) p(T ))MT @T + (v(i) p(i) ) @Q (T v(T ) p(T ))MT @T

T 2U

(6.6.40)

where v(i) , p(i) is the solution to the problem (v(i) p(i) ) 2 Xk Mk

Here, @Q and @ (@Qp) are de ned by (6.6.22). @T u Proof. By virtue of (6.6.6), (6.6.39), and Theorem 6.6.2, the functionals i are continuously Frechet di erentiable in N1 and their derivatives are determined by
i (T )MT =
0

@ i @Q (i) (i) @ (u p) (T v(T ) p(T )) (v p ) = ; @ (v p) (T v(T ) p(T )):

(6.6.41)

Since we have

@ i (T v(T ) p(T ))MT @T + @@v ip) (T v(T ) p(T )) F 0 (T )MT (

T 2 U:

(6.6.42)

@ i (T v(T ) p(T )) 2 (X M ) k k @ ( v p)

F 0 (T )MT 2 Xk Mk

@ i (T v(T ) p(T )) F 0 (T )MT @ (v p) i = @@v p) (T v(T ) p(T )) F 0(T )MT = A: ( A = ; @ (@Qp) (T v(T ) p(T )) (v(i) p(i) ) F 0 (T )MT u = (v(i) p(i) ) @Q (T v(T ) p(T ))MT : @T

(6.6.43)

From (6.6.37), (6.6.41), and (6.6.43) (6.6.44)

Now, (6.6.40) follows from (6.6.42) and (6.6.44).

476

Chapter 6. Optimization problems for steady ows

:::

be a solution to the problem (6.6.8) with v(T ), p(T ) de ned by (6.6.34){(6.6.36) for an arbitrary xed k. Then, there exist constants i not all equal to zero such that
m+4 X i i=0 i 0i (T0 )(T

Theorem 6.6.4 Suppose the conditions of Theorem 6.6.3 are satis ed, and let T0

; T0 ) 0

T 2N

(6.6.45) (6.6.46) (6.6.47)

0 i i (T0 ) = 0 ^ where T = (F F ~), T0 = (F0

m+1 (T ) = kF k2 2 (S2 )n ; c2 1 L (T ) = k kp pl (S3 ) ; cp m+3 3 W

i = 0 1 ::: r m +1 ::: m+4 i = 1 ::: r m+1 ::: m +4

^ F0
0 ~0 )

and

^ L m+2 (T ) = kF k2 2 (S3 )n ; c2 2 ~kp l (;) ; cp m+4 (T ) = k Wp 4

(6.6.48)

c1 , : : : , c4 being the constants from (6.6.7).

Proof. By using the notations (6.6.48), the set Nad from (6.6.7) can be represented
in the form Nad = T j T 2 N i (T ) 0 i = 1 : : : r m + 1 : : : m + 4 i (T ) = 0 i = r + 1 : : : m r m : (6.6.49)

It follows from (6.6.4) and (6.6.18) that N is a convex set in U , N1 is an open set in U , and N N1 , see (6.6.19). By Theorem 6.6.3, the functionals i , i = 0 1 : : : m, are continuously Frechet di erentiable in N1 , and so are the functionals m+1 : : : m+4 via Theorem 1.10.1. Finally, by applying Theorem 1.12.2, we obtain (6.6.45){(6.6.47).

6.7 Problems on the optimal shape of a canal

We suppose that the uid ows in a canal that occupies the following domain G, which is represented in cylindrical coordinates (r z ), see Fig. 6.7.1,

G = (r

z) j 0 < z < l 0

< 2 0 r < ( z) :

(6.7.1)

The boundary ; of the canal is de ned by ;1 = (r ;2 = (r ;3 = (r ; = ;1 ;2 ;3 z) j r = ( z) 0 <2 0<z<l z ) j z = 0 0 r ( 0) 0 <2 z ) j z = l 0 r ( l) 0 <2 : (6.7.2)

6.7. Problems on the optimal shape of a canal

x2

477

h a ( , z)

0
a

x3 , z

x1

Figure 6.7.1: Domain occupied by a canal and disposition of cylindrical and Cartesian coordinate system The cross-sections z = 0 and z = l, i.e., ;2 and ;3 , are the inlet and outlet of the canal. The following problem often appears in practice. We are given the above cross-sections, and, in addition, a function F of surface forces at the inlet may be given. The optimization problem consists in nding the length l of the canal, its lateral area, i.e., the function , and the function of surface forces F (if it is not given) with a view to optimize, in some sense, the ow of the uid. In what follows, we will consider these problems.

6.7.1 Set of controls and di eomorphisms

We denote by E the transformation of cylindrical coordinates into Cartesian ones:

E : (r z ) ! E (r z ) = (x1 x2 x3 ) x1 = r cos x2 = r sin x3 = z: The transformation E is a bijection of the set W1 = (r z ) j (r z ) 2 R3 0 < r < 1 0 ' < 2 z 2 R
onto the complement of the straight line

(6.7.3) (6.7.4)

A = x j x = (x1 x2 x3 ) 2 R3 x1 = x2 = 0 z 2 R
and it is a di eomorphism of the C 1 class of the set

W = (r

z ) j (r

z ) 2 R3 r 2 (0 1)

2 (0 2 ) z 2 R

(6.7.5)

478

Chapter 6. Optimization problems for steady ows

:::

onto E (W ). By E ;1 we denote the inverse of E . Let also N = ( z ) j 2 (0 2 ) z 2 (0 1) : We determine a set of controls of the form
n

(6.7.6)

(6.7.7) Here, r1 , l1 are given positive constants, 1 and 2 given functions. The set M is equipped with the topology generated by that of C 1 (N ) R. To each q = (f l) 2 M, we assign the domain q 2 R3 of the form
q=
n

@f @f M = q j q = (f l) f 2 C 1 (N ) f (0 z ) = f (2 z ) @ (0 z ) = @ (2 z ) z 2 0 1] f ( z ) r1 ( z ) 2 N f ( 0) = 1 ( ) o f ( 1) = 2 ( ) 2 0 2 ] l 2 R l l1 :

x j x = (x1 x2 x3 ) 0 < x3 < l

By B we denote the cylinder of unit length and radius , i.e., B = x j x = (x1 x2 x3 ) 2 R3 0 x2 + x2 < 0 < x3 < 1 1 2 and let = Br1 r2 2 R 0 < r2 < r1 : To each q 2 M, we assign a mapping Pq;1 : ! q by if x 2 Br2 Pq;1 (x) = (x1 x2 x3 l) ;1 (E A(q) E )(x) if x 2 n Br2 : Here
(

0 (x2 + x2 ) 1 < f ( xl3 ) 1 2 2

x o = arctan x2 : (6.7.8)
1

(6.7.9) (6.7.10) (6.7.11)

A(q): E ;1 ( n Br2 ) ! Aq (E ;1 ( n Br2 )) (r z ) ! A(q)(r z ) = ( (q r z ) zl)

(q r

z) =

2 X

where

i=0

ai (q

z )ri

(6.7.12)

a2 (q a1 (q a0 (q

z z ) = f((r ;)r;)r1 1 2 2 z ) = 1 ; 2r2 a2 (q z ) 2 z ) = f ( z ) ; r1 a1 (q z ) ; r1 a2 (q

(6.7.13)

z ):

6.7. Problems on the optimal shape of a canal

479

Lemma 6.7.1 Let a set M be de ned by (6.7.7) and let domains q and be determined by (6.7.8) with (f l) = q 2 M and (6.7.9), (6.7.10). Then, for each q 2 M, the mapping Pq;1 given by (6.7.11){(6.7.13) is a C 1 di eomorphism of onto q , and its inverse Pq is a C 1 di eomorphism of q onto . Proof. Applying (6.7.12) and (6.7.13) (see also subsec. 2.12.4), one can easily verify that (q r2 z ) = r2 (q r1 z ) = f ( z ) (6.7.14) @ (q r z ) = 1 ( z) 2 N : 2
. Moreover,

@r

It follows from (6.7.7) and (6.7.13) that a2 (q Therefore,

@ (q r @r

z ) = 1 + 2(r ; r2 )a2 (q

z ):

(6.7.15)

@ (q r @r

z ) 0. Thus,

z) 1

q 2 M r 2 r2 r1 ] ( z ) 2 N :

det(A(q))0 = @ l l in E ;1 ( n Br2 ) @r and there exists a mapping Pq that is the inverse of Pq;1 . Since Pq;1 is a continuously di erentiable bijection of onto q , we obtain the inclusion Pq 2 C 1 ( q ) from the theorem on the di erentiability of an implicit function. Remark 6.7.1 If f is a smooth function, one can construct, in a way similar to that in subsec. 2.12.4, a C m di eomorphism of onto q , where m 2. For q = (f l) 2 M, we denote by Sq the boundary of q . It follows from (6.7.8) that Sq = Sq1 Sq2 Sq3 , where

6.7.2 Optimization problems

Here, = arctan x2 , 1 , 2 are the functions from (6.7.7). x1 We consider the following problem on ow of a nonlinear viscous uid: Find a pair vq , pq satisfying

Sq1 = x j x = (x1 x2 x3 ) 2 R3 (x2 + x2 ) 1 = f ( xl3 ) 0 < x3 < l 1 2 2 Sq2 = x j x = (x1 x2 x3 ) 2 R3 x3 = 0 0 (x2 + x2 ) 1 1 ( ) 1 2 2 3 x = l 0 (x2 + x2 ) 1 Sq3 = x j x = (x1 x2 x3 ) 2 R 3 2( ) : 1 2 2

(6.7.16)

@pq ; 2 @ ('(I (v ))" (v )) = 0 q ij q @xi @xj

in q i = 1 2 3

(6.7.17)

480

Chapter 6. Optimization problems for steady ows

:::

div vq = 0

; pq ij + 2'(I (vq ))"ij (vq ) ; pq ij + 2'(I (vq ))"ij (vq )

vq Sq1 = 0

in q

qj Sq2 = Fi ^ qj Sq3 = Fi

i=1 2 3 i = 1 2 3:

(6.7.18) (6.7.19) (6.7.20) (6.7.21)

^ Here, qj are the components of the unit outward normal q to Sq , and Fi , Fi the ^ of surface forces at the inlet and outlet of components of the functions F and F the canal. We note that when q passes through M, the set Sq3 moves along the x3 axis without deformations. So, we consider Sq2 and Sq3 as xed sets in R2 that are independent of q. We de ne the following spaces

Xq = u j u 2 H 1 ( q )3 u Sq1 = 0 Vq = u j u 2 Xq div u = 0 X = u j u 2 H 1 ( )3 u S 1 = 0
where

(6.7.22) (6.7.23) (6.7.24)

S1 = x j x = (x1 x2 x3 ) 2 R3 (x2 + x2 ) 1 = r1 x3 2 (0 1) : 1 2 2
Suppose the function ' satis es the conditions (6.1.11){(6.1.14) and de ne operators Lq : Xq ! Xq and Bq 2 L(Xq L2 ( q ) ) by (Lq (u) h) = 2 (Bq u ) =
Z Z
q q

'(I (u))"ij (u)"ij (h) dx u 2 Xq

u h 2 Xq

(6.7.25) (6.7.26)

(div u) dx

2 L2( q ):

^ Assume that F is a xed function and 1 ^ F 2 H ; 2 (Sq3 )3

(6.7.27)

^ while F is a control. If the uid ows out into the air, we can take F = (0 0 p0), where p0 is a constant, the atmospheric pressure. We assume that U = M 1 H ; 2 (Sq2 )3 is a set of controls. For (q F ) 2 U , we consider the following problem: Find a pair of functions vqF , pqF satisfying (vqF pqF ) 2 Xq L2 ( q ) ^ (Lq (vqF ) h) ; (Bq pqF h) = (F + F h) h 2 Xq (Bq vqF ) = 0 2 L2 ( q ): (6.7.28) (6.7.29) (6.7.30)

6.7. Problems on the optimal shape of a canal

481

The pair vq = vqF , pq = pqF is a generalized solution to the problem (6.7.17){ (6.7.21). The existence and uniqueness of the solution to the problem (6.7.28){ (6.7.30) follow from Theorem 6.1.1. We de ne a goal functional in the form
0 (q

F) =

3 X

where yi are prescribed functions from L2(Sq3 ) and vqFi are the components of the vector function vqF . In particular, for the extrusion head, one can take y1 = y2 = 0, y3 = c, where c is a constant. A set of admissible controls is determined as follows f Uad = (q F ) j q = (f l) 2 M F 2 L2 (Sq2 )3 f 2 Wt2 (N ) kf kWt2(N ) c1 t > 2 kF kL2(Sq2 )3 c2 l 2 l1 c3 ] : (6.7.32) f
f Here, Wt2 (N ) is the subspace of Wt2 (N ) consisting of periodic functions with respect to , c1 , c2 , c3 are positive constants, l1 is the constant from (6.7.7). The optimization problem consists in nding a pair q0 , F0 satisfying (q0 F0 ) 2 Uad (6.7.33) 0 (q0 F0 ) = (q Finf U 0 (q F ): )2
ad

Sq3 i=1

(vqFi ; yi )2 ds

(6.7.31)

Theorem 6.7.1 Suppose a function ' satis es the conditions (6.1.11){(6.1.14) and (6.7.27) holds. Let a nonempty set Uad and a goal functional 0 be de ned by (6.7.32) and (6.7.31) with yi 2 L2 (Sq3 ). Then, there exists a solution of the problem (6.7.33). Proof. Let vqF = vqF Pq;1, pqF = pqF Pq;1, where vqF , pqF is the solution to ~ ~ the problem (6.7.28){(6.7.30) and Pq;1 is de ned by (6.7.11){(6.7.13). By virtue of Lemma 6.7.1, the pair vqF , pqF is the solution to the problem ~ ~ (~qF pqF ) 2 X L2 ( ) v ~ (6.7.34) e q (~qF ) h) ; (Bq pqF h) = (Q1 (q F ) + Q2 (q F ) h) e ~ ^ (L v h2X (6.7.35) e ~ (Bq vqF ) = 0 2 L2( ): (6.7.36)
e e Here, the operators Lq and Bq are de ned by (6.3.22) and (6.3.23), and ^ ^ (Q1 (q F ) h) = (F h Pq ) (Q2 (q F ) h) = (F h Pq ): (6.7.37) As Sq2 and Sq3 are xed sets in R2 , the functions Q1 and Q2 are independent 2 of q. So, from the conditions that qm ! q0 in M, Fm ! F0 in H ; 1 (Sq2 )3 , it follows that Q2 (qm Fm ) ! Q2 (q0 F0 ) in X . In addition, by (6.7.11){(6.7.13), we get Pq;1 ! Pq;1 in C 1 ( )3 . These results and Lemma 6.7.1 allow us to apply m 0 Theorem 6.3.1, which gives ) (q F ) ! (~qF pqF ) is a continuous mapping of v ~ (6.7.38) 2 M H ; 1 (Sq2 )3 into X L2 ( ),

482

Chapter 6. Optimization problems for steady ows

:::

2 and so 0 is a continuous functional in M H ; 1 (Sq2 )3 . Because of the embedding 1 ; 2 (S )3 . Hence, the existence of a solution theorem, Uad is a compactum in M H q2 to the problem (6.7.33) follows from Theorem 1.3.9. The proof is complete. We now consider another optimization problem. Given cross-sections at the inlet and outlet of a canal and functions of surface forces at the inlet and outlet, nd the length of the canal and its lateral area which maximize the volume ow rate of the uid. So, in this case, for each q 2 M, the functions of velocity and pressure vq , pq are the solution to the problem

(vq pq ) 2 Xq L2 ( q ) ^ (Lq (vq ) h) ; (Bq pq h) = (F + F h) h 2 Xq (Bq vq ) = 0 2 L2( q ): ^ Here, F and F are given xed functions, 1 ^ 2 F 2 H ; 2 (Sq3 )3 F 2 H ; 1 (Sq2 )3 : The goal functional is de ned by
0 (q ) =

(6.7.39) (6.7.40) (6.7.41) (6.7.42) (6.7.43)

Sq 3

vqi qi ds

where vqi are the components of the velocity function vq , qi the components of the unit outward normal q to Sq3 . In our case, q = (0 0 1) on Sq3 . The set of admissible controls has the form
f Uad = q j q = (f l) 2 M f 2 Wt2 (N ) kf kWt2 (N ) c1 t > 2 l 2 l1 c2 ] : f

(6.7.44)

The optimization problem consists in nding q0 satisfying

q0 2 Uad

0 (q0 ) =

q2Uad

sup

0 (q ):

(6.7.45)

By analogy with the proved above, we obtain

Theorem 6.7.2 Let a function ' satisfy the conditions (6.1.11){(6.1.14) and let (6.7.42) hold. Let vq , pq be the solution to the problem (6.7.39){(6.7.41), and 0 be de ned by (6.7.43). Let also a nonempty set Uad be de ned by (6.7.44). Then, there exists a solution to the problem (6.7.45).
We note that the problem on the shape optimization of a two-dimensional canal with a view of the maximization of the volume ow rate of a viscous Newtonian uid was studied in Litvinov (1990a).

6.8. A problem on the optimal shape of a hydrofoil

483

6.8.1 State equation for a moving hydrofoil

Boundary-value problem

6.8 A problem on the optimal shape of a hydrofoil

Let a nondeformable hydrofoil move slowly in a viscous incompressible uid. We suppose that the dimensions of the cross-sections of the hydrofoil are small with respect to its length, and the cross-sections vary slowly in the longitudinal direction. Therefore, we consider the two-dimensional problem of the ow about a hydrofoil of a constant cross-section. We place the origin of the coordinate system in some internal point of the hydrofoil and assume that the coordinate system is rigidly attached to the hydrofoil and moves with it. In polar coordinate system, the boundary Sq of the hydrofoil is determined by a smooth periodical function q : 0 2 ] ! R, and we designate by Qq the domain occupied by the hydrofoil (see Fig. 6.8.1). Let Q be a bounded and su ciently large domain in R2 containing the domain Qq . In polar coordinate system, the boundary S of Q is de ned by a smooth periodical function : 0 2 ] ! R. Denote q = Q n Qq and consider the following problem: Find a pair (v p), where v = (v1 v2 ) is the velocity and p is the pressure, satisfying the relations

@ 2 v1 + @ 2 v1 ; @p = 0 in q 2 2 @y1 @y2 @y1 @ 2 v2 + @ 2 v2 ; @p = 0 in q 2 2 @y1 @y2 @y2 @v1 + @v2 = 0 in q @y1 @y2 v Sq = 0 v S = a: a(y) = (a1 (y) a2 (y)) = c = (c1 c2 ) 2 R2 y 2 S:

(6.8.1) (6.8.2) (6.8.3) (6.8.4)

Here, is viscosity of the uid and a = (a1 a2 ), where the functions ai take constant values, that is, (6.8.5) Clearly, ;c = (;c1 ;c2 ) is the velocity of points of the hydrofoil in the immovable uid. We assume that the ratio jcj is large therefore, we use the Stokes approximation for our problem. Such an approximation is suitable for the case when jcj is small. For large c, the ow is turbulent, but the equations (6.8.1), (6.8.2) may be used if is considred as the turbulent viscosity. This approach traces back to Boussinesq (1877). The turbulent viscosity is far greater than the laminar one (it may be greater than the laminar viscosity by a factor 103{105), see Loitzansky (1987), Fletcher (1988). So, the ratio jcj may be large for large jcj if is the turbulent viscosity. For other models of turbulent ows, see Rodi (1980), Litvinov (1996).

484

Chapter 6. Optimization problems for steady ows

y2

:::

Qq W

a
Sq
q

q(a )

g a( )

y1

Qq

Figure 6.8.1: Domain Qq occupied by a hydrofoil and a large domain Q containg

The problem (6.8.1){(6.8.4) is considered in a Holder space C l ( q ), where l > 0, l not an integer. The space C l ( q ) is equipped with the norm (2.13.3). We denote by C l ( q )=R the quotient space of C l ( q ) with respect to R. In the other domains, the spaces C l are de ned analogously. We consider the function q, determining the shape of the hydrofoil, as a control. De ne a space of controls M of the form
e M = q j q 2 C l]+4 ( 0 2 ]) r1 < q( ) < r2

Set of controls and operator Aq

202 ] :

(6.8.6)

e Here, r1 , r2 are positive constants, l > 0, l not an integer, C l]+4 ( 0 2 ]) is the l]+4 ( 0 2 ]). We equip the set M with the subspace of periodical functions in C topology generated by that of C l]+4 ( 0 2 ]). For each q 2 M , we de ne in polar coordinate system the boundary Sq between the domain Qq occupied by the hydrofoil and the domain q = Q n Qq see Fig. 6.8.1. We assume below that the domain Q is independent of the control q. De ne spaces Vlq and Hlq as follows:

Vlq = C l+2 ( q )2 C l+1 ( q ) Hlq = C l ( q )2 C l+1 ( q ) C l+2 (Sq )2 C l+2 (S )2 :

Also, determine an operator Lq by the formulas

(6.8.7)

Lq = (Aq Bq ) 2 L(Vlq Hlq ) (u p) 2 Vlq Aq (u p) = Mu ; grad p div u Bq (u p) = u Sq u S :

(6.8.8)

6.8. A problem on the optimal shape of a hydrofoil

485

Let f , , , be functions such that f = (f1 f2) 2 C l ( q )2 = ( 1 2 ) 2 C l+2 (Sq )2 =( and

Z
q

2 C l+1 ( q ) l+2 2 2 ) 2 C (S )
Z

(6.8.9) (6.8.10)

dy =

where qi and i are the components of the unit outward normals q and to Sq and S . In (6.8.10) and in the following formulas, summation over repeated index is implied. The theorem below follows from known results, see Solonnikov (1964, 1966). Theorem 6.8.1 Let q be a domain in R2 with two connected components of the boundary. The external boundary S and internal boundary Sq of q are de ned e in polar coordinate system by the functions 2 C l]+4 ( 0 2 ]) and q 2 M . Let functions f , , , satisfy the conditions (6.8.9){(6.8.10). Then, there exists a unique pair (~ p) 2 C l+2 ( q )2 C l+1 ( q )=R satisfying v~ Aq (~ p) = (f ) v~ Bq (~ p) = ( ) v~ (6.8.11) and there exists a positive constant c such that ; kvkC l+2 ( q )2 + kpkC l+1( q )=R c kf kC l( q )2 + k kC l+1( q ) ~ ~ (6.8.12) + k kC l+2(Sq )2 + k kC l+2(S)2 :
In particular, there exists a unique pair (v p) 2 C l+2 ( q )2 C l+1 ( q )=R satisfying the equations (6.8.1){(6.8.4).

Sq

i qi ds +

i i ds

Operator Wq and functionals

^ We denote by Vlq and Hlq the kernel subspace and the range of the operator Lq = (Aq Bq ), ^ Vlq = (u p) j (u p) 2 Vlq Lq (u p) = 0 Hlq = Lq (Vlq ): (6.8.13) ^ ^ We denote by Vlq and Hlq the complements of Vlq and Hlq in Vlq and Hlq , that is, ^ Vlq = Vlq Vlq ^ Hlq = Hlq Hlq

(6.8.14) where the symbol represents the direct sum of subspaces. From Theorem 6.8.1, ^ it follows that Vlq is a one-dimensional subspace in Vlq and that ^ Vlq = (u p) j u = 0 p 2 R : (6.8.15)

486

Chapter 6. Optimization problems for steady ows

:::

^ From Theorem 6.8.1 and (6.8.10), it also follows that Hlq is the one-dimensional subspace in Hlq determined by the expression ^ Hlq = (f ) 2 Hlq f = 0 = c = ;c q = ;c c 2 R : (6.8.16) From (6.8.15){(6.8.16), it follows that the function ' = (0 1), with 0 the zero ^ element of C l+2 ( q )2 and 1 the unit function in q , is a basis in Vlq , while the ^ function = (0 1 ; q ; ) is a basis in Hlq . Taking into account Theorem 6.8.1, it is easy to obtain the following e Theorem 6.8.2 Let q 2 M and 2 C l]+4( 0 2 ]). Determine the operator Uq 2 L(Vlq Hlq ) of the form Uq ' = Uq w = 0 w 2 Vlq : (6.8.17) Then, the operator Wq = Lq + Uq is an isomorphism of Vlq onto Hlq . We further denote by (vq pq ), with vq = (vq1 vq2 ), the solution of the problem (6.8.1){(6.8.4). The hydrodynamic lift F1 (q) of the hydrofoil occupying the domain Qq is de ned by the formula

F1 (q) =

Sq

; pq 2j +
Z

@vq2 + @vqj @yj @y2

qj ds

(6.8.18)

where ij is the Kroneker delta. Because

Sq q2 ds = 0

the integral (6.8.18) is uniquely de ned for pq 2 C l+1 ( q )=R. The dissipation energy F0 (q) is de ned by the formula

F0 (q) = 2

2 X
q i j =1

@vqi + @vqj @yj @yi

dy

(6.8.19)

and this energy is equal to the power (rate of work) needed to move the hydrofoil. We denote by E the function that transforms polar coordinates (r ) in Cartesian coordinates (y1 y2 ), E : (r ) ! E (r ) = (y1 y2) y1 = r cos y2 = r sin (6.8.20) and the inverse of E is denoted by E ;1 . We denote by q1 the domain corresponding to q in polar coordinates (r ) (see Fig. 6.8.1), (6.8.21) q1 = (r ) j 2 (0 2 ) q ( ) < r < ( ) :

State equation in polar coordinates

6.8. A problem on the optimal shape of a hydrofoil

487

We de ne sets Sq1 and S1 as follows:

Sq1 = (r ) j 2 (0 2 ) r = q( ) S1 = (r ) j 2 (0 2 ) r = ( ) :

(6.8.22)

These sets correspond to the internal and external boundaries of q in polar coordinates. Let u = (u1 u2 ) be a velocity eld in q . The function w = (w1 w2 ), de ned in q1 by the formulas

w1 (r ) = u1 (E (r )) cos + u2(E (r )) sin w2 (r ) = ;u1 (E (r )) sin + u2 (E (r )) cos (6.8.23) corresponds to the u in polar coordinate system, with w1 , w2 being the radial and tangential components of the velocity vector w. To a function of the pressure p in q , we place in correspondence the function p1 = p E in q1 . Then, with the
replacement of the variables and functions de ned by (6.8.20) and (6.8.23), the operator Lq = (Aq Bq ) from (6.8.8) transforms into the operator Lq1 of the form

Lq1 = (Aq1 Bq1 ) 2 @w2 w1 @p1 M1 w1 ; 2 r @ ; r2 ; @r Aq1 (w p1 ) = > M1 w2 + 22 @w1 ; w22 ; 1 @p1 r @ r r @ > > @w > > 1 + 1 @w2 + w1 : @r r @ r Bq1 (w p1 ) = w Sq1 w S1 :
8 > > > > > <

(6.8.24) (6.8.25)

Here, M1 is the Laplace operator in polar coordinates, @2 + 1 @ + 1 @2 : M =

1

function satis es the conditions

e 2 C l]+4 ( 0 2 ])

(6.8.26) r @r r2 @ 2 In the sequel, we suppose that q 2 M , where M is de ned by (6.8.6) and the

@r2

' 2 (0 2 ) (6.8.27) where 0 is a constant and r2 is the constant from (6.8.6). By q and ~, we denote ~ the periodical extensions of the functions q and on R with the period (0 2 ),
(')
0

> r2

and we de ne the set

q 2 = (r ) j

2 R q( ) < r < ~( ) : ~

~ For k > 0, k not an integer, we denote by C k ( q1 ) the subspace of functions in C k ( q1 ) periodical with respect to . The periodicity of a function h 2 C k ( q1 )

488

Chapter 6. Optimization problems for steady ows

:::

means that, if ~ is periodical with respect to the -extension of h on q2 with h ~ period (0 2 ), then h 2 C k (G), where G is an arbitrary, open, and bounded subset of q2 . De ne functions F1 and F2 as follows:

F1 : (0 2 ) ! Sq1 ! F1 ( ) = (q ( ) ) F2 : (0 2 ) ! S1 ! F2 ( ) = ( ( ) ): (6.8.28) If u is a function given on Sq1 , then u F1 is a function de ned on (0 2 ), and we determine the space C k (Sq1 ), where k 2 0 l] + 4], of the form C k (Sq1 ) = u j u F1 2 C k ( 0 2 ]) (6.8.29a)
and by analogy

C k (S1 ) = u j u F2 2 C k ( 0 2 ]) :

(6.8.29b)

e e By C k (Sq1 ) and C k (S1 ), we denote the subspaces of periodical functions in C k (Sq1 ) k (S1 ). In the spaces C k (Sq1 ) and C k (S1 ), the norms are de ned by e e and C

kukC k (Sq1 ) = ku F1 kC k ( 0 2 e

])

kukC k (S1 ) = ku F2 kC k ( 0 2 ]) : e

(6.8.30)

To the spaces Vlq and Hlq from (6.8.7), we place in correspondence the following spaces:
e e Vlq1 = C l+2 ( q1 )2 C l+1 ( q1 ) e e e e Hlq1 = C l ( q1 )2 C l+1 ( q1 ) C l+2 (Sq1 )2 C l+2 (S1 )2 : (6.8.31) Let g, e, h, z be functions such that e e g = (g1 g2 ) 2 C l ( q1 )2 e 2 C l+1 ( q1 ) e e h = (h1 h2 ) 2 C l+2 (Sq1 )2 z = (z1 z2 ) 2 C l+2 (S1 )2 : (6.8.32) Consider the following problem: Find a pair (w b) satisfying the relations Lq1 (w b) = (g e h z ) (w b) 2 Vlq1 : (6.8.33)

From (6.8.10) and Theorem 6.8.1, it follows that the condition of solvability of the problem (6.8.32){(6.8.33) has the form
Z
q1

er dr d =

Z2

0 Z2 0

hi (q( ) ) q1i ( ) (q0 ( ))2 + (q( ))2 2 d zi ( ( ) ) 1i ( ) ( (

0

))2 + (

1 ))2 2

d :

(6.8.34)

6.8. A problem on the optimal shape of a hydrofoil

489

Here, q0 , 0 are the derivatives of q, , and q1i , 1i are the radial (i = 1) and angular (i = 2) components of the unit outward normals to Sq and S . These components are de ned by the formulas 0 0 q11 = ; cos(arctan( qq )) q12 = sin(arctan( qq )) 0 0 (6.8.35) 12 = ; sin(arctan( )): 11 = cos(arctan( )) By analogy with (6.8.14), we have the following representation: ^ ^ Vlq1 = Vlq1 Vlq1 Hlq1 = Hlq1 Hlq1 (6.8.36) where ^ Vlq1 = ker Lq1 Hlq1 = Lq1 (Vlq1 ) ^ and Vlq1 , Hlq1 are the corresponding complements. From Theorem 6.8.1 and ^ ^ (6.8.34), it follows that Vlq1 and Hlq1 are one-dimensional subspaces. The function 'q , 'q = (0 1) (6.8.37) ^ e is the basis in Vlq1 , with 0 the zero element of C l+2 ( q1 )2 and 1 the unit function in q1 . The function q , (6.8.38) q = 0 1 (; q1i )2=1 (; 1i )2=1 i i ^ is the basis in Hlq1 . De ne an operator Uq1 2 L(Vlq1 Hlq1 ) as follows:

Uq1 'q = q

Uq1 w = 0

w 2 Vlq1 :

(6.8.39)

The operator Uq1 is determined by the formulas

e e u 2 C l+2 ( q1 )2 p 2 C l+1 ( q1 ) Uq1 (u p) = c(q) (q) Z
;1 Z

c(q) =

Remark 6.8.1 By (6.8.6) and (6.8.27), the function

in the case when q or to Hlq1 . From Theorem 6.8.1, we get the following theorem, which is analogous to Theorem 6.8.2. Theorem 6.8.3 Let the condition (6.8.27) hold and let q 2 M . Let operators Lq1 = (Aq1 Bq1 ) and Uq1 be de ned by (6.8.24), (6.8.25), (6.8.40). Then, Lq1 is an isomorphism of Vlq1 onto Hlq1 and Wq1 = Lq1 + Uq1 is an isomorphism of Vlq1 onto Hlq1 .

q belongs to Hlq1 . However, e l]+3 ( 0 2 ]), the function q may not belong are from C

q1

r dr d

q1

pr dr d :

(6.8.40)

490

Chapter 6. Optimization problems for steady ows

:::

State equations in a xed domain

We de ne a domain by

6.8.2 Fixed-domain problem and Frechet di erentiability of the functionals

= x j x = (x1 x2 ) 2 R2 1 < x2 + x2 < 4 : 1 2 ( ') j 2 (1 2) ' 2 (0 2 ) ( ') j = 1 ' 2 (0 2 ) ( ') j = 2 ' 2 (0 2 ) : (6.8.41)

We denote by 1 the domain corresponding to in polar coordinates ( ') and by S2 , S3 the sets corresponding to the internal and external boundaries of ,
1= S2 = S3 =

(6.8.42)

To each q 2 M , we assign the mapping Gq : q1 ! 1 of the form (r ) ! Gq (r ) = ( ') 2 )+ ( = r ; ( q( ; q( ) ) '= : ) The inverse G;1 : q
1 ! q1

(6.8.43)

acts as follows: = ': (6.8.44)

( ') ! G;1 ( ') = (r ) q r = 2q(') ; (') + (') ; q(')] We de ne the following spaces:

e e Vl1 = C l+2 ( 1 )2 C l+1 ( 1 ) l ( 1 )2 C l+1 ( 1 ) C l+2 (S2 )2 C l+2 (S3 )2 e e e e Hl1 = C

(6.8.45)

e which correspond to the spaces Vlq1 , Hlq1 from (6.8.31). The norms in C k (Si ), i = 2 3, k 2 0 l] + 4], are de ned by formulas analogous to (6.8.30). By applying the composite function theorem (chain rule), it is easy to prove the lemma below.

Lemma 6.8.1 Let (6.8.27) hold, and let a function Gq be de ned by (6.8.43). Then, Gq is a di eomorphism of q1 onto 1 of the C l]+4 class, and the mapping e e e e u ! u Gq is an isomorphism of C k ( 1 ) onto C k ( q1 ) and of C k (S2 ), C k (S3 ) e e onto C k (Sq1 ), C k (S1 ), respectively, for k 2 0 l] + 4].
sions For every q 2 M , we de ne operators Lq2 , Uq2 2 L(Vl1 Hl1 ) by the expres-

Lq2 = (Aq2 Bq2 )

6.8. A problem on the optimal shape of a hydrofoil

491

Aq2 h = (Aq1 (h Gq )) G;1 q Bq2 h = (Bq1 (h Gq )) G;1 q ;1 Uq2 h = (Uq1 (h Gq )) Gq @ @r @' @r

Let
e w = (w1 w2 ) 2 C l+2 ( q1 )2 u = (u1 u2 ) = w G;1 q

h 2 Vl1 :
1 @r G;1 = ; ; q @' q G;1 = 1: q
e p1 2 C l+1 ( q1 ) p = p1 G;1 : q

(6.8.46)

By (6.8.43), (6.8.44), we obtain the following formulas: 1 G;1 = ; q q G;1 = 0 q

@ @ @' @

(6.8.47)

e e From Lemma 6.8.1, it follows that u 2 C l+2 ( 1 )2 , p 2 C l+1 ( 1 ). We take h = ((u1 u2 ) p) in (6.8.46). By applying (6.8.24), (6.8.25), (6.8.47) and the composite function theorem, we obtain the following formulas for the operators Aq2 , Bq2 :

Aq21 (u p) = Aq2 (u p) = : Aq22 (u p) (u p) 2 Vl1 Aq23 (u p) ; Bq2 (u p) = u S2 u S3

where
2 2 Aq21 (u p) = a11 (q) @@ u21 + a22 (q) @ u21 @' @ 2 u1 + b (q) @u1 + b (q) @u2 + a12 (q) @ @' 1 @ 2 @ @u2 + b (q)u ; b (q) @p ; b3 (q) @' 4 1 5 @ 2 2 Aq22 (u p) = a11 (q) @@ u22 + a22 (q) @ u22 @' @ 2 u2 + b (q) @u2 ; b (q) @u1 + a12 (q) @ @' 1 @ 2 @ @u1 + b (q)u + b (q) @p ; b (q) @p + b3 (q) @' 4 2 6 @ 7 @' Aq23 (u p) = b5 (q) @u1 ; b6 (q) @u2 + b7 (q) @u2 + u1 : @ @ @'

8 <

(6.8.48) (6.8.49)

(6.8.50)

492

Chapter 6. Optimization problems for steady ows

2

:::

Here, we use the following notations

Note that r is de ned by (6.8.44), from which one obtains the partial derivatives of r, which are present in (6.8.51). From (6.8.40) and (6.8.46), it follows that the operator Uq2 is de ned by

@r a11 (q) = ( ; q);2 1 + r12 @' a22 (q) = r12 @r a12 (q) = ; r2 ( 2; q) @' 2r @r 1 b1 (q) = r12 @' ( ; q)2 @@ @' + 0 ; q0 @2 1 + ; q r ; @'r 2 1 @r b2 (q) = r22 ; q @' b3 (q) = r22 b4 (q) = ;r;2 b5 (q) = ( ; q);1 1 @r b6 (q) = 1 ; q @' b7 (q) = r;1 : r

(6.8.51)

Uq2 (u p) = c(q) q G;1 (u p) 2 Vl1 q Z ;1 Z @r @r pr @ d d' c(q) = r @ d d' 1 1

(6.8.52)

where q is determined by (6.8.35), (6.8.38). By applying Theorem 6.8.3 and Lemma 6.8.1, we obtain the following theorem. Theorem 6.8.4 Let (6.8.27) hold and let M be de ned by (6.8.6). Then, for each q 2 M , the operator Lq2 + Uq2 , where Lq2 = (Aq2 Bq2 ) and Uq2 are de ned by (6.8.48){(6.8.52), is an isomorphism of Vl1 onto Hl1 . By virtue of Theorem 6.8.4, for each q 2 M there exists a unique pair (uq gq ) satisfying the relations Lq2 (uq gq ) = (0 0 0 a) Uq2 (uq gq ) = 0 (uq gq ) 2 Vl1 (6.8.53) where a is the vector function from (6.8.4) taking constant values see (6.8.5). De ne a pair (v p) as follows: v = (N1 (uq Gq )) E ;1 p = gq Gq E ;1 (6.8.54) where N1 is the following matrix: sin N1 = cos ; cos : (6.8.55) sin

6.8. A problem on the optimal shape of a hydrofoil

493

It easy to see that the pair (v p) de ned by (6.8.54) is a solution of the problem (6.8.1){(6.8.4) and (v p) 2 Vlq . By Theorem 6.8.4, there exists a unique solution to the problem (6.8.53) for an arbitrary q 2 M . So, we de ne a mapping N : M ! Vl1 such that

N (q) = (uq gq )

q2M

(6.8.56)

where (uq gq ) is the solution of the problem (6.8.53). Below, we study some properties of the mapping N . We now establish three lemmas, by which we prove the di erentiability of the mapping N .

Continuity and Frechet di erentiability of the mapping N

Lemma 6.8.2 Let 2 = (r1 r2 ) R 1, where r1 , r2 are the positive constants from (6.8.6) and let a function f : (x1 x2 ') ! f (x1 x2 ') be twice continuously di erentiable in 2 . Let the function f have arbitrary continuous derivatives up to and including order four, for which the orders of the derivatives with respect to ' do not exceed two, and these derivatives may be extended by continuity onto r1 r2 ] R 1 . Assume that X is a bounded, open, and convex set in M , where M is given by (6.8.6) for l 2 (0 1), and M , X are equipped with the topology generated by the C 4 ( 0 2 ]) topology. De ne mappings f1 and f2 from X into C 2 ( 1 ) as follows :
f1 (q)( ') = f (q(') q0 (') ') f2 (q)( ') = f (q(') q00 (') ') q2X
Then, f1 and f2 are continuously Frechet di erentiable mappings from X into C 2 ( 1 ), and the Frechet derivatives f10 , f20 of these mappings at a point q 2 X are de ned by the formulas

@f 0 (f1 (q)h)( ') = @x (q(') q0 (')

1 2

')h(') ')h0 (') ')h(') ')h00 (')

e h 2 C 4 ( 0 2 ]): (6.8.57)

@f 0 (f2 (q)h)( ') = @x (q(') q00 (')

1 2

@f + @x (q(') q0 (') @f + @x (q(') q00 (')

Proof. By applying the Taylor formula and by virtue of simple but cumbersome
estimates, we infer that

kfi(q + h) ; fi (q) ; fi0 (q)hkC (

1)

c1 khk2 2( 0 2 C

])

(6.8.58)

494

Chapter 6. Optimization problems for steady ows

:::

@k 2 0 (6.8.59) @ k1 @'k;k1 fi (q + h) ; fi (q) ; fi (q)h C ( 1 ) c2 khkC 4( 0 2 ]) here, i = 1 2 k = 1 2 0 k1 k. Also, q, q + h 2 X , fi0 (q) are de ned by (6.8.57) and the constants c1 , c2 depend on f , X and do not depend on h. Let now fqn g X , q 2 X , and qn ! q in C 4 ( 0 2 ]). Then, the following estimate holds
lim n = 0 i = 1 2: Therefore, the function q ! fi0 (q) is a continuous mapping of X into
e L(C 4 ( 0 2 ]) C 2 ( 1 )):

k(fi0 (qn ) ; fi0 (q))hkC 2 (

1)

n khkC 4 ( 0 2 ])

e h 2 C 4 ( 0 2 ])

By analogy with the proof of Lemma 6.8.2, we can prove the lemma below.

Lemma 6.8.3 Let X be an arbitrary, bounded, open, and convex subset of M , where M is de ned by (6.8.6) with l 2 (0 1). Let also 3 = (r1 r2 ) R, where r1 , r2 are the constants from (6.8.6), and let f be a function that is continuous together with the derivatives Dk f , jkj 5, in 3 , and these derivatives allow a e continuous extension onto r1 r2 ] R. De ne a function f1 : X ! C 3 ( 0 2 ]) by f1 (q)(') = f (q(') q0 (')) q 2 M:
e Then, f1 is a continuously Frechet di erentiable mapping of X into C 3 ( 0 2 ]), 0 of this function at a point q 2 X is given by and the Frechet derivative f1

(6.8.60) @f (q(') q0 ('))h0 (') e 4 ( 0 2 ]): + @x h2C 2 Lemma 6.8.4 Let a set M be de ned by (6.8.6) and equipped with the topology generated by the topology of C l]+4 ( 0 2 ]). Let operators Aq2 , Bq2 , Uq2 be dened by (6.8.48){(6.8.52), and let (6.8.27) hold. Then, the functions q ! Lq2 = (Aq2 Bq2 ), q ! Uq2 are continuous mappings from M into L(Vl1 Hl1 ) for an arbitrary l > 0, l not an integer. For l 2 (0 1), these functions are continuously
Frechet di erentiable.

@f 0 (f1 (q)h)(') = @x (q(') q0 ('))h(')

1

Proof. It is easily seen that

for an arbitrary k > 0, the multiplication of functions u v !> = e e uv is a bilinear continuous mapping from C k ( 1 ) C k ( 1 ) > e into C k ( 1 ).

(6.8.61)

6.8. A problem on the optimal shape of a hydrofoil

495

e Let now fqn g M , q 2 M and qn ! q in C l]+4 ( 0 2 ]). From (6.8.44), (6.8.51), (6.8.61), it follows that

aij (qn ) ! aij (q) bi (qn ) ! bi (q) bi (qn ) ! bi (q) Aqn 2i ! Aq2i Aqn 23 ! Aq23

e in C l]+1 ( 1 ) e in C l]+1 ( 1 ) e in C l]+2 ( 1 )

i j=1 2 i=1 2 3 4 i = 5 6 7:

(6.8.62)

Further, from (6.8.50) after taking into account (6.8.61) and (6.8.62), we obtain
e C l ( 1 )) i=1 2 el+1 ( 1 )): (6.8.63) C From (6.8.49), it follows that the operator Bq2 is independent of q therefore, Bqn 2 = Bq2 , and (6.8.63) yields that q ! Lq2 = (Aq2 Bq2 ) is a continuous mapping of M into L(Vl1 Hl1 ). By applying (6.8.35), (6.8.38), (6.8.44), (6.8.52), we infer that Uqn 2 ! Uq2 in L(Vl1 Hl1 ). Lemma 6.8.2 and the formulas (6.8.51) imply that 9 for l 2 (0 1), the functions q ! aij (q), with i j = 1 2, and> = the functions q ! bi (q), with i = 1 : : : 7, are continuously (6.8.64) > 2 ( 1 ). e Frechet di erentiable mappings of M into C From (6.8.48){(6.8.50), (6.8.61), (6.8.64) it follows that, for l 2 (0 1), the function q ! Lq2 = (Aq2 Bq2 ) is a continuously Frechet di erentiable mapping from M into L(Vl1 Hl1 ). By analogy, applying Lemma 6.8.3 and (6.8.35), (6.8.38), (6.8.44), (6.8.52), we obtain that, for l 2 (0 1), the function q ! Uq2 is a continuously Frechet di erentiable mapping from M into L(Vl1 Hl1 ), concluding the proof. Theorem 6.8.5 Let a set M be de ned by (6.8.6) and equipped with the topology generated by the topology of C l]+4 ( 0 2 ]) let operators Aq2 , Bq2 , Uq2 be de ned by (6.8.48){(6.8.52) let Lq2 = (Aq2 Bq2 ) and (6.8.27) hold. Then, the function N : q ! N (q) = (uq gq ), where uq , gq is the solution of the problem (6.8.53), is a continuous mapping of M into Vl1 and for l 2 (0 1), the function N is a continuously Frechet di erentiable mapping of M into Vl1 and the Frechet derivative N 0 of N at a point q 2 M is given by e N 0 (q)h = ;(Lq2 + Uq2 );1 ((L0q2 + Uq0 2 )h)(uq gq ) h 2 C 4 ( 0 2 ]): (6.8.65) Here, (Lq2 + Uq2 );1 is the inverse of Lq2 + Uq2 and L0q2 , Uq0 2 are the Frechet derivatives of the functions q ! Lq2 , q ! Uq2 at a point q. Proof. De ne a mapping J : M Vl1 ! Hl1 as follows: q2M (u g) 2 Vl1 (6.8.66) J (q (u g)) = (Lq2 + Uq2 )(u g) ; (0 0 0 a):

in L(Vl1 in L(Vl1

496

Chapter 6. Optimization problems for steady ows

:::

Here, a is the vector function from (6.8.4) taking constant value. It is obvious that the function N : M ! Vl1 introduced in (6.8.56), where uq , gq is the solution of the problem (6.8.53), is an implicit function de ned by the mapping J , i.e.,

J (q N (q)) = 0:

(6.8.67)

The existence of the function N follows from Theorem 6.8.4. To prove the continuity of N , we use the implicit function theorem (Thorem 1.9.1). By Lemma 6.8.4, we infer that J is a continuous mapping from M Vl1 into Hl1 . For an arbitrary xed q 2 M , the function J (q ): (u g) ! J (q (u g)) is an a ne continuous mapping from Vl1 onto Hl1 , and the Frechet derivative of it has the form

@J @ (u g) (q (u g)) = Lq2 + Uq2 :

(6.8.68)

From here and Lemma 6.8.4, we obtain that (q (u g)) ! @ (@Jg) (q (u g)) is a conu tinuous mapping from M Vl1 into L(Vl1 Hl1 ). From Theorem 6.8.4 and (6.8.68), it follows that @ (@Jg) (q (u g)) is an isomorphism of Vl1 onto Hl1 . By the implicit u function theorem, we now infer that N is a a continuous mapping from M into Vl1 . Taking Lemma 6.8.4 into account, one can easily see that, in case l 2 (0 1), the function J is a continuously Frechet di erentiable mapping from M Vl1 into Hl1 . From the theorem on di erentiability of an implicit function (Thorem 1.9.2), we now obtain that, for l 2 (0 1), the function N is a continuously Frechet di erentiable mapping from M into Vl1 , and its Frechet derivative is de ned by (6.8.65).

Continuity and di erentiability of the functionals F0 and F1

We now change in (6.8.18), (6.8.19) the functions vq = (vq1 vq2 ), pq and the variables y1, y2 for the functions uq = (uq1 uq2 ), gq and the variables , '. Taking notice of (6.8.20), (6.8.44), (6.8.54) for v = vq , p = pq , we obtain the following formulas for the functionals F1 and F0 :

F1 (q) =

Z2
0

;gq 1i + 2 "1i(uq )]

q1i sin '

=1

+ ;gq 2i + 2 "2i (uq )] q1i cos '

Z 2Z 2 X 2
1 0

(q0 )2 + q2

1 2

d'

(6.8.69) (6.8.70)

F0 (q) = 2

i j =1

@r ("ij (uq ))2 r @ d d':

6.8. A problem on the optimal shape of a hydrofoil

497

Here, q1i and r are de ned by (6.8.35), (6.8.44), and "ij (uq ) are determined by the formulas 1 "11 (uq ) = ; q @uq1 @ @r "22 (uq ) = ; r( 1 q) @' @uq2 + 1 @uq2 + uq1 ; @ r @' (6.8.71) "12 (uq ) = "21 (uq ) @r 1 q = 1 ; r( 1 q) @' @uq1 + 1 @uq1 + ; q @uq2 ; ur 2 : 2 ; @ r @' @ We note that "ij (uq ) Gq are the components of the rate of strain tensor in polar coordinates r, (i.e., in the domain q1 ) for the velocity eld uq Gq . e De ne mappings eij : M C l+2 ( 1 )2 ! C 1 ( 1 ), where i j = 1 2, as follows: e q2M u = (u1 u2) 2 C l+2 ( 1 )2 e (q u) = 1 @u1
11

From (6.8.71) and (6.8.72), it follows that eij (q uq ) = "ij (uq ): (6.8.73) We now de ne functions f1 : M Vl1 ! C ( 0 2 ]) and f2 : M Vl1 ! C ( 1 ) as follows: q2M ( u g ) 2 Vl 1 f1 (q (u g)) = ;g 1i + 2 e1i (q u)] q1i sin ' 1 + ;g 2i + 2 e2i (q u)] q1i cos ' =1 (q0 )2 + q2 2

@r e22 (q u) = ; r( 1 q) @' @u2 + 1 @u2 + u1 ; @ r @' e12 (q u) = e21 (q u) @r 1 1 = 2 ; r( 1 q) @' @u1 + 1 @u1 + ; q @u2 ; u2 : ; @ r @' @ r

;q @

(6.8.72)

from (6.8.69), (6.8.70) take now the forms

2 X @r (eij (q u))2 r @ : (6.8.74) f2 (q (u g)) = 2 i j =1 Here, q1i and r are determined by (6.8.35), (6.8.44). The functionals F1 and F0

F1 (q) =

f1 (q (uq gq )) d' 0 Z 2Z 2 F0 (q) = f2 (q (uq gq )) d 1 0

Z2

d':

(6.8.75)

498

Chapter 6. Optimization problems for steady ows

:::

Theorem 6.8.6 Let the conditions of Theorem 6.8.5 hold, and let functionals F1 and F0 be given by (6.8.69), (6.8.70). Then, for arbitrary l > 0, l not an integer, the functionals F1 and F0 are continuously Frechet di erentiable in M , and their Frechet derivatives F10 and F00 at a point q 2 M are de ned as follows :
@f1 (q (u g ))h + @f1 (q (u g )) N 0 (q) h d' (6.8.76) q q q q @q @ (u g) 0 Z 2Z 2 @f2 (q (u g ))h + @f2 (q (u g )) N 0 (q) h d d' F00 (q)h = q q q q @q @ (u g) 1 0 F10 (q)h =
(6.8.77)
e where h 2 C l]+4 ( 0 2 ]), N 0 (q)h is given by (6.8.65), and @fi , @ (@fig) are the partial @q u Frechet derivatives of the functions fi , i = 1 2, with respect to the rst and second arguments. Z2

Proof. By analogy with the above (see the proofs of Lemmas 6.8.2 and 6.8.4), applying (6.8.35), (6.8.44), (6.8.72), and (6.8.74), we establish that f1 , f2 are continuously Frechet di erentiable mappings from M Vl1 into C ( 0 2 ]) and C ( 1 ). Further, the functions
u!
Z2
0

u d'

u!

Z 2Z 2
1 0

u d d'

are linear continuous mappings from C ( 0 2 ]) and C ( 1 ) into R. Now, for l 2 (0 1), Theorem 6.8.6 and the formulas (6.8.76), (6.8.77) follow from the composite function theorem and Theorem 6.8.5. In this case, directly from the de nition of Frechet derivative, it follows that, for l 2 (0 1) and for arbitrary q and q + h from M , the following estimates hold:

jF1 (q + h) ; F1 (q) ; F10 (q)hj (h)khkC 4( 0 2 ]) jF0 (q + h) ; F0 (q) ; F00 (q)hj (h)khkC 4( 0 2 ]) where (h) ! 0 as h = 0, khkC 4( 0 2 ]) ! 0. If now l > 1 and h ! 0 in M , then all 6 the more khkC 4( 0 2 ]) ! 0. Hence, Theorem 6.8.6 is valid for an arbitrary l > 0, l
not an integer.

6.8.3 Optimization problem

On the set M , we de ne a functional F2 as follows: 1 Z 2 q2 d': F2 (q) = 2 0 (6.8.78)

6.8. A problem on the optimal shape of a hydrofoil

499

It is obvious that F2 (q) is the area of a hydrofoil occupying the domain Qq (see Fig. 6.8.1). It is easily seen that the functional F2 is continuously Frechet di erentiable in M and its Frechet derivative F20 at a point q 2 M is given by

F20 (q)h =

Z2
0

qh d'

e h 2 C l]+4 ( 0 2 ]):

(6.8.79)

We de ne a set of admissible controls in the form e M1 = q j q 2 C l+4 ( 0 2 ]) kqkC l+4( 0 2 ]) c e (6.8.80) r3 q(') r4 ' 2 (0 2 ) F1 (q) c1 F2 (q) = c2 : Here, l > 0, l not an integer, c, c1 , c2 , r3 , r4 are positive con-9 = stants, and r1 < r3 < r4 < r2 , where r1 , r2 are the positive (6.8.81) constants from (6.8.6). Consider the following optimization problem: Find a function q0 satisfying the relations

F0 (q0 ) = qmin F0 (q) 2M

1

q0 2 M1 :

(6.8.82)

From the physical viewpoint, the problem (6.8.82) consists in nding the shape of the hydrofoil (the function q0 ) which would require minimal power to move the hydrofoil, while the constraints on the area and on the hydrodynamic lift are satis ed.

Theorem 6.8.7 Let functionals F0 , F1, F2 be de ned by (6.8.18), (6.8.19), (6.8.78), where pq = p, vq = v = (v1 v2 ), and (p v) is the solution of the problem (6.8.1){(6.8.4). Let also (6.8.27) hold, and let a nonempty set M1 be de ned by (6.8.80), (6.8.81). Then, there exists a solution of the problem (6.8.82).
satisfying

Proof. As M1 is a nonempty set, there exists a minimizing sequence fqk g1 k=1

lim F0 (qk ) = q2M F0 (q) inf
1

fqk g M1 :

(6.8.83)

e e As the embedding of C l+4 ( 0 2 ]) into C l]+4 ( 0 2 ]) is compact, it is possible to extract a subsequence, still denoted by fqk g, such that

qk ! q0
From here, by Theorem 6.8.6, we have

e in C l]+4 ( 0 2 ]):

(6.8.84) (6.8.85)

F0 (qk ) ! F0 (q0 )

F1 (qk ) ! F1 (q0 ):

500

Chapter 6. Optimization problems for steady ows

:::

By (6.8.84), we also have F2 (qk ) ! F2 (q0 ), and as fqk g M1 by (6.8.84) we obtain kq0 kC l+4( 0 2 ]) c (6.8.86) e where c is the constant from (6.8.80). Now, by applying (6.8.83){(6.8.86), we conclude that the function q0 is a solution of the problem (6.8.82). We now establish the necessary optimality conditions. De ne a set M2 by e M2 = q j q 2 C l+4 ( 0 2 ]) kqkC l+4( 0 2 ]) c r3 q(') r4 ' 2 (0 2 ) : e (6.8.87)
e We equip the set M2 with the topology generated by the topology of C l+4 ( 0 2 ]). By (6.8.80), we have M1 = q j q 2 M2 F1 (q) c1 F2 (q) = c2 : (6.8.88) Theorem 6.8.8 Let the conditions of Theorem 6.8.7 hold and let a function q0 be a solution of the problem (6.8.82). Then, there exists constants 0 , 1 , 2 , not all equal to zero, such that
2 X

0 1 0 If q0 is an internal point of M2, then

0 2 X

i=0

i Fi0 (q0 ) (q ; q0 )

q 2 M2

1 (F1 (q0 ) ; c1 ) = 0:

(6.8.89)

Proof. M2 is a convex subset of M , and F0 , F1 , F2 are continuously Frechet di erentiable functionals in M . Now, Theorem 6.8.8 follows from (6.8.88) and Theorem 1.12.2.

i=0

i Fi0 (q0 ) = 0:

6.9 Direct and optimization problems with consideration for the inertia forces
We have above studied optimization problems for steady ows under neglect of the inertia forces. Now, we consider the general problem on steady ows of a nonlinear viscous uid taking into account these forces. We will show that, for non-high-speed ows of high-viscous uids, there exists a unique solution of the direct problem, and the above results on the optimization of uid ows may be transferred to the case of consideration for the inertia forces. It should be noted that the majority of real nonlinear viscous uids are highviscous, and they usually ow slowly.

6.9. Direct and optimization problems with consideration for the inertia forces 501

6.9.1 Setting and solution of the direct problem

Equations and the de nition of a generalized solution
We consider the problem on steady ows of the nonlinear viscous uids with the constitutive equation (6.1.1) in a bounded domain Rn , n = 2 or 3, with a Lipschitz continuous boundary S . By (6.1.1), (6.1.4), (6.1.5), we have the following equations of motion and incompressibility:

@ '(I (v))"ij (v) @v @p vj @xi + @x ; 2 = Ki @xj j i n X @v div v = @xi = 0 i=1 i

in in :

i = 1 ::: n

(6.9.1) (6.9.2)

Let S1 , S2 , S3 be open subsets of S such that S = S 1 S 2 S 3 , Si \ Sj = ? for i 6= j . We consider the mixed boundary conditions (6.1.7){(6.1.10). The spaces X and V are the same as in (6.1.19) and (6.1.20), i.e.,

X = u j u 2 H 1 ( )n u S1 = 0 (u ; u ) S3 = 0 V = u j u 2 X div u = 0

(6.9.3) (6.9.4)

and the norm in X and V is given by (6.1.21). We suppose that the functions of ^ volume forces K and surface forces F , F acting on S2 and S3 satisfy the conditions ^ K2X F 2X F 2X : (6.9.5) Suppose also that the functions ' and satisfy the conditions (6.1.11){(6.1.14) and (6.1.15){(6.1.18). Let us de ne operators L : X ! X , N : X ! X , and B 2 L(X L2 ( )) as follows

'(I (u))"ij (u)"ij (h) dx + (juj2 )ui hi ds S3 Z @u (N (u) h) = uj @x i hi dx j Bu = div u:

(L(u) h) = 2 Let also

(6.9.6) (6.9.7) (6.9.8) (6.9.9)

M = L + N:
(v p) 2 X L2 ( )

A generalized solution of the problem (6.9.1), (6.9.2), (6.1.7){(6.1.10) is dened to be a pair v, p satisfying (6.9.10)

502

Chapter 6. Optimization problems for steady ows

:::

^ (M (v) h) ; (B p h) = (K + F + F h) h2X (Bv q) = 0 q 2 L2 ( ):

(6.9.11) (6.9.12)

By using Green's formula, one may show that, if v, p is a solution to the problem (6.9.10){(6.9.12), then v, p is a solution to the problem (6.9.1), (6.9.2), (6.1.7){(6.1.10) in the distribution sense. On the contrary, if v, p is a solution to the problem (6.9.1), (6.9.2), (6.1.7){(6.1.10) that satis es (6.9.10), then v, p is a solution to (6.9.10){(6.9.12).

and using the Holder inequality and the embedding theorem, we obtain (N (uk ) ; N (u0 ) h)
Z

The existence and uniqueness Lemma 6.9.1 Suppose that n = 2 or 3 and uk ! u0 weakly in X : Then, N (uk ) ! N (u0 ) strongly in X : Proof. Denoting Z Ak = u0j @uki ; @u0i hi dx @x @x
j j j

(6.9.13)

ckuk ; u0kL4 ( )n kuk kX khkX

The norm in the space Lq ( )n is de ned by

(ukj ; u0j ) @uki hi dx + jAk j @x

+ jAk j:

(6.9.14)

kukLq (
Z

)n

Z X n

i=1

1 q dx q jui j

q 2 1 1):

(6.9.15)

Taking (6.9.3) into account and applying again Green's formula, the Holder inequality, and the embedding theorem, we get

jAk j =

@ (u (u ; u )h ) @xj oj ki 0i i @h ; u0j (uki ; u0i ) @x i ; @u0j (uki ; u0i )hi dx @x

Z

S2 S3 n X i j =1

u0j (uki ; uoi )hi j ds + c1 ku0 kX kuk ; u0kL4 ( )n khkX

S3 ) k;(uki ; u0i )kL2 (S2 S3 ) k;hi kL4 (S2 S3 )
)n khkX :

k;u0j kL4 (S2

+ c1 ku0kX kuk ; u0 kL4(

(6.9.16)

6.9. Direct and optimization problems with consideration for the inertia forces 503

Here, ; is the trace operator, ;u = u S2 S3 . It follows from (6.9.13) that uk ! u0 in L4( )n and ;uk ! ;u0 in L2(S2 S3 )n . So, (6.9.14) and (6.9.16) give N (uk ) ! N (u0 ) in X , which completes the proof. De ne positive constants and as follows: = =
u2V kukX 1 u2V kukX 1

sup

sup

^ j(K + F + F u)j:
2

@u uj @x i ui dx
j

(6.9.17) (6.9.18) (6.9.19)

p

Lemma 6.9.2 Suppose the conditions (6.1.11){(6.1.18) are satis ed and

< a1
where a1 is the positive constant from (6.1.12). Then, the following estimate holds :

= : (6.9.20) Proof. By applying (6.1.12), (6.1.16), (6.9.9), (6.9.17), and (6.9.18), we get ^ (M (u) u) ; (K + F + F u) f (kukX ) = 2a1 kuk2 ; kuk3 ; kukX u 2 V: (6.9.21) X X Consider the quadratic equation 2a1y ; y2 ; = 0: (6.9.22) Its roots are those of the equation f (y) = 0 and they are equal to p p a1 + a2 ; : a1 ; a2 ; 1 1 y = y = (6.9.23)
1 2

^ (M (u) u) ; (K + F + F u) 0

2 if u 2 V and kukX = a1 ; a1 ;

If (6.9.19) holds, then y1 and y2 are real and f (y) 0 for y 2 y1 y2 ]. So, (6.9.19) yields (6.9.20), and the lemma is proved. For l > 0, we denote by d(0 l) the closed ball in V of radius l centered at 0, i.e., d(0 l) = u j u 2 V kukX l : (6.9.24) Let a be a trilinear form in X generated by the operator N :

a(u v w) =
and
1

@v uj @xi wi dx
j u v2d(0 1)

(6.9.25) (6.9.26)

= sup

u v2d(0 1)

a(u v v)

2=

sup

a(v u v) :

504

Chapter 6. Optimization problems for steady ows

:::

Lemma 6.9.3 Let the conditions (6.1.11){(6.1.18) be satis ed, let (6.9.19) hold,
and let be de ned by (6.9.20). Let also

1 ; ( 1 + 2)

>0

(6.9.27)

where 1 is the positive constant from (6.1.46). Then, the operator M is strictly monotone in d(0 ), i.e.,

u w 2 d(0 ) (6.9.28) Proof. Let u, w be arbitrary functions from V and h = u ; w. From (6.9.7) and
(6.9.26), we obtain (N (u) ; N (w) h) =
Z ;

(M (u) ; M (w) u ; w)

ku ; wk2 X

2 1 kukX + 2 kwkX khkX :

@h uj @x i + @wi hj hi dx @x
j j

(6.9.29)

Now, (6.1.46), (6.9.9), and (6.9.29) yield (6.9.28).

Theorem 6.9.1 Let the conditions (6.1.11){(6.1.18) be satis ed and let (6.9.19) hold. Then, there exists a solution v, p of the problem (6.9.10){(6.9.12) such that kvkX . If, additionally, (6.9.27) holds, then there exists a unique solution of (6.9.10){(6.9.12) such that v 2 d(0 ): Proof. It follows from (6.9.10){(6.9.12) that the function v is a solution of the
problem

v2V ^ (M (v) h) = (K + F + F h)
k!1 z2Vk

h 2 V: u2V

(6.9.30) (6.9.31) (6.9.32) (6.9.33)

Let fVk g be a sequence of nite-dimensional subspaces of V such that lim inf ku ; z kX = 0

Vk Vk+1 : We search for the Galerkin approximations vk satisfying vk 2 Vk ^ (M (vk ) h) = (K + F + F h) h 2 Vk :

By virtue of Lemma 6.9.2, there exists a solution of the problem (6.9.33) and kvk kX . Thus, we can extract a subsequence fvm g such that vm ! v0 weakly in V . We pass to the limit as m ! 1 in (6.9.33), with k changed by m. In this case, we use the monotonicity of the operator L and the compactness of the operator N , Lemma 6.9.1. Thus, the function v = v0 is a solution of the problem (6.9.30).

6.9. Direct and optimization problems with consideration for the inertia forces 505

that

From Lemma 6.1.1, it follows that there exists a function p 2 L2( ) such ^ B p = M (v) ; K ; F ; F: (6.9.34)

So, the pair v, p is a solution to the problem (6.9.10){(6.9.12). In the case when (6.9.27) holds, the operator M is strictly monotone in d(0 ), see Lemma 6.9.3. Therefore, in the ball d(0 ), there exist a unique solution of the problem (6.9.30) and a unique solution of the problem (6.9.10){(6.9.12) such that v 2 d(0 ).

Remark 6.9.1 Theorem 6.9.1 does not state that, under the conditions (6.1.11){ (6.1.18), (6.9.19), and (6.9.27), there exists a unique solution of the problems (6.9.30) and (6.9.10){(6.9.12). But a unique solution such that v 2 d(0 ) does
exist. Let fXk g and fMk g be sequences of nite-dimensional subspaces of X and L2 ( ) which satisfy the conditions (6.1.37), (6.1.38), and (6.1.42). We search for an approximate solution of the problem (6.9.10){(6.9.12) in the form (vk pk ) 2 Xk Mk ^ (M (vk ) h) ; (Bk pk h) = (K + F + F h) h 2 Xk (Bk vk q) = 0 q 2 Mk : (6.9.35) (6.9.36) (6.9.37)

6.9.2 Approximation of the problem (6.9.10){(6.9.12)

From the point of view of applications, in particular, of computation, the problem (6.9.35){(6.9.37) is considerably more preferable than (6.9.33). So, we study the question on convergence of the approximations vk , pk . De ne constants 1 , 3 , 1 , 4 , 5 by
k k p 2 a1 ; a1 ; 1 3 1= 4 = sup umax a(u 1 k v2dk 5 = sup umax a(v u v ) k v2dk k

= sup max a(u u u) u2d

k

^ = sup max (K + F + F u) u 2d

v v)
(6.9.38) (6.9.39)

where a1 is the constant from (6.1.12) and

dk = u j u 2 Xk (Bk u ) = 0
Note that
1

2 Mk kukX 1 :
2,

1, 5

see (6.9.17), (6.9.18), (6.9.26).

Theorem 6.9.2 Let the conditions (6.1.11){(6.1.18) be satis ed and let fXk g, fMk g be sequences of nite-dimensional subspaces of X and L2 ( ) that satisfy

506

Chapter 6. Optimization problems for steady ows

:::

(6.1.37), (6.1.38), (6.1.42) and Xk also

Xk+1 , Mk
3<

Mk+1 for an arbitrary k. Let

(6.9.40)

a2 : 1
1

Then, for each k, there exists a solution of the problem (6.9.35){(6.9.37), and from the sequence fvk pk g one can extract a subsequence fvm pm g such that vm ! v in X , pm ! p in L2 ( ), where v, p is a solution of the problem (6.9.10){(6.9.12). If, additionally,
1 ; ( 4 + 5) 1

>0

(6.9.41)

then for each k there exists a unique solution of the problem (6.9.35){(6.9.37) such that kvk kX 1 :

sketch of it. By using (6.9.40) and (6.1.42), we argue the solvability of the problem (6.9.35){(6.9.37) for each k, and the boundedness of the solutions. By selecting appropriate subsequences, we pass to the limit, using the monotonicity of the operator L and the compactness of the operator N , see Lemma 6.9.1. If (6.9.41) holds, then for each k the operator L = M + N is strictly monotone in the ball

Proof of this theorem is analogous to that of Theorem 6.1.1, so we give only a

dk 1 = u j u 2 Xk (Bk u q) = 0 q 2 Mk kukX 1 : (6.9.42) In this case, kvk kX 1 and there exists a unique solution of the problem (6.9.35){ (6.9.37) such that vk 2 dk 1 .

6.9.3 Some remarks on models, optimization problems, and existence results

For small data, more exactly, for the case when (6.9.19) holds, we have established above the results on the existence of a solution of the general problem on steady ows of the nonlinear viscous uid with the mixed boundary conditions (6.1.7){ (6.1.10), when zero velocities, a function of surface forces, and the conditions of ltration are prescribed on di erent parts of the boundary. If, additionally, (6.9.27) holds, a unique solution of this problem exists in a small ball. It should be mentioned that, for large data, i.e., for large velocities, our problem, as well as the corresponding problem for the Navier{Stokes equations, which are obtained from (6.9.1) when ' = const, is ill-posed. Moreover, the equations for the nonlinear viscous uid and the Navier{Stokes equations do not describe the main phenomena observed at high-speed and turbulent ows of real uids, see Litvinov (1996), and they are not suitable for such ows. This is why we only considered ows with small velocities, which are well described by our equations.

6.9. Direct and optimization problems with consideration for the inertia forces 507

Notice also that not only the equations (6.9.1), but also those (6.1.6) may be used for such ows. The optimization problems for the Navier{Stokes equations without assumptions on the smallness of velocities and with a function of volume forces considered as a control were rst set and studied by Fursikov (1980, 1982, 1983a). For further investigation of optimization problems for the Navier{Stokes equations, see Abergel and Temam (1990), Sritharan (1991, 1992), Gunzburger et al. (1991, 1992), Casas (1993). By using the continuity and strict monotonicity of the operator M in a small ball of the space V , for small data, one can carry over, by analogy, all the results of Sections 6.2 and 6.3 on the continuity of the function control-solution of the direct problem to the case of consideration for the inertia forces, provided that the viscosity of the uid is large, while the velocities are not large.

508

Chapter 6. Optimization problems for steady ows

:::

Bibliography
Abeles, F. (1950). Recherches sur la propagation des ondes electromagnetiques sinusoidales dans les milieux strati es. Application aux couches minces. Annales de Physique, Ser. 12, 5, 596{782 Abergel, F. and Temam, R. (1990). On some control problems in uid mechanics. Theoret. Comput. Fluid Dynamics. 1, 303{325 Adams, R.A. (1975). Sobolev Spaces. Academic Press. New York Agmon, S., Douglis, A., and Nirenberg, L. (1959). Estimates near the boundary for solutions of elliptic partial di erential equations satisfying general boundary condtions I. Com. Pure Appl. Math. 12, 623{727 Agmon, S., Douglis, A., and Nirenberg, L. (1964). Estimates near the boundary for solutions of elliptic partial di erential equations satisfying general boundary conditions II. Com. Pure Appl. Math. 17, 35{92 Alberg, J.H., Nilson, E.N., and Walsh, J.L. (1967). The Theory of Splines and Their Applications. Academic Press. New York Allaire, G. (1994). Explicit lamination parameters for three-dimensional shape optmization. Control and Cybernetics 23, 309{326 Allaire, G. and Kohn, R. (1993). Optimal bounds on the e ective behavior of a mixture of two well-ordered elastic materials. Quart. Appl. Math. 51, 643{671 Ambarstumian, S.M. (1974). General Theory of Anisotropic Shells. Nauka. Moscow (in Russian) Armand J.-L. (1972). Applications of the Theory of Optimal Control of DistributedParameter Systems to Structural Optimization. NASA Report CR-2044 Armand, J.-L. and Lodier, B. (1978). Optimal design of bending elements. Int. J. Num. Meth. Engng. 13, 373{384 Astarita, G. and Marrucci, G. (1974). Principles of Non-Newtonian Fluid Mechanics. McGraw-Hill. London 509

510

Bibliography

Aubin, J.-P. (1972). Approximation of Elliptic Boundary Value Problems. Wiley{ Interscience. New York Baiocchi, C. and Capelo, A. (1984). Variational and Quasivariational Inequalities |Applications to Free Boundary Problems. John Wiley and Sons. New York Bakhvalov, N.S. (1980). Homogenization and perturbation problems. In Computing Methods in Applied Science and Engineering, pp. 645{658. North-Holland. Amsterdam Bakhvalov N.S. and Panasenko, G.P. (1984). Homogenization of Processes in Periodic Media. Nauka. Moscow (in Russian) Banerjee, P.K. and Butter eld, R. (1981). Boundary Element Methods in Engineering Scince. McGraw-Hill. London Banichuk, N.V. (1980). Optimization of the Shapes of Elastic Bodies. Nauka. Moscow (in Russian) English transl.: Problems and Methods of Optimal Structural Design. Plenum Press. 1983 Banichuk, N.V., Kartvelishvili, V.M., and Mironov, A.A. (1977). Numerical solutions of two-dimensional optimization problems for elastic plates. Mechanics of Solids 12, 65{74 Bends e, M.P. (1994). Methods for optimization of structural topology, shape and material. Preprint, Mathematical Institute, Technical University of Denmark Bends e, M. and Guedes, J. (1994). Some computational aspects of using extremal material properties in the optimal design of shape, topology and material. Control and Cybernetics 23, 327{349 Bends e, M. and Kikuchi, N. (1988). Generating optimal topologies in structural design using a homogenizatiion method. Comp. Meth. Appl. Mech. Engrg. 71, 197{224 Bends e, M.P, and Rodrigues, H.C. (1990). On topology and boundary variations in shape optimization. Control and Cybernetics 19, no. 3{4, 9{26 Bensoussan, A., Lions, J.L., and Papanicolau, G. (1978). Asymtotic analysis for Periodic Structures. North Holland. Amsterdam Bernadou, M., Palma, F.J., and Rousselet, B. (1991). Shape optimization of an elastic thin shell under various criteria. Structural Optimization 3, 7{21 Besov, O.V., Il'in, V.P., and Nikol'skii, S.M. (1975). Integral Representation of Functions and Embedding Theorems. Nauka. Moscow (in Russian) English transl.: Vol. 1, 2, Wiley, 1979 Bolotin, V.V. and Novikov, Yu.N. (1980). Mechanics of Multilayer Constructions.

Bibliography

511

Mashinostroenie. Moscow (in Russian) Born, M. and Wolf, E. (1964). Principles of Optics. Pergman Press. Oxford Boussinesq, J. (1877). Essai sur la theorie des eaux courantes. In Memoires Presentees par Diverses Savants a l'Acad. d. Sci., Vol. 23. Paris Bourbaki, N. (1955). Elements de Mathematique, Livre V, Espaces Vectoriels Topologiques. Hermann. Paris Bourbaki, N. (1960). Elements de Mathematique, Livre III, Topologie Generale. Hermann. Paris Bratus, A.S. (1981). Asymptotic solutions in problems of optimal control by coefcients of elliptic operators. DAN SSSR 259, 1035{1038 (in Russian) Bratus, A.S. and Seiranian, A.P. (1983a). Two multiple eigenvalues in optimization problems. DAN SSSR 272, 275{278 (in Russian) Bratus, A.S. and Seiranian, A.P. (1983b). Bimodal solutions in problems of optimization of eigenvalues. Prikl. Matem. Mekh. 47, 546{554 (in Russian) Brezis, H. (1983). Analyse Fonctionnelle, Theorie et Applications. Masson. Paris English transl.: Springer-Verlag. Heidelberg, 1987 Burak, Ya.I. and Budz, S.F. (1974). Determination of the optimal rate of heating of a thin spherical shell. Prikl. Mekh. 10, no. 2, 14{20 (in Russian) Burak, Ya.I. and Domanskii, P.P. (1982). Optimization of dynamic e ects in shells of revolution under axisymmetric force load. Soviet Appl. Mechan. 18, no. 2, 92{98 Burak, Ya.I., Zozuliak, Yu.D., and Gera, B.V. (1984). Optimization of Transients for Thermoelastic Shells. Naukova Dumka. Kiev (in Russian) Burczynski, T. and Fedelinski, P. (1990). Shape sensitivity analysis and optimal design of static and vibrating systems using the boundary element method. Control and Cybernetics 19, no. 3{4, 47{71 Calderon, A.P. (1961). Lebesgue spaces of di erentiable functions and distributions. In Proc. Sympos. Pure Math. Vol. 4, pp. 33{49. Providence Casas, E. (1990). Optimality conditions and numerial approximations for some optimal design problems. Control and Cybernatics 19, no. 3{4, 73{91 Casas, E. (1993). Some optimal control problems of turbulent ows. In Optimal Design and Control of Structures under Statical and Vibtration Response. Advanced Tempus Course. Lecture Notes, Part 1, pp. 1{17, Banach Centre, Poland Cea, J. (1971). Optimisation. Theorie et Algorithmes. Dunod. Paris

512

Bibliography

Cea, J. (1978). Numerical search method for an optimal domain. In Numerical Methods in Mathematical Physics, Geophysics, and Optimal Control (Sympos. Novosibirsk (1976), Marchuk, G.I. and Lions, J.L., eds., pp. 64{74. Nauka. Novosibirsk (in Russian) Cea, J. and Malanowski, K. (1970). An example of a max-min problem in partial di erential equations. SIAM J. Control 8, 305{316 Chenais, D. (1987). Optimal design of midsurface of shells: di erentiability and sensitivity computation. Appl. Math. and Optimization 16, 93{133 Chenais, D. (1994). Shape optimization of shells., Control and Cybernetics 23, 351{382 Cherkaev, A. V. (1994). Relaxation of problems of optimal structural design. Int. J. Solids and Structures 31, 2251{2280 Christensen, R.M. (1979). Mechanics of Composite Materials. Wiley-Interscience. New York Chudinovich, I.Yu. (1991). Methods of Boundary Equations in Problems of the Elastic Medium Dynamics. Kharkov University. Kharkov (in Russian) Ciarlet, P. (1978). The Finite Element Method for Elliptic Problems. North-Holland. Amsterdam Donnell, L.H. (1976). Beams, Plates, and Shells. McGraw-Hill. New York Dunford, N. and Schwartz, J.T. (1958). Linear Operators. Part I: General Theory. Interscience Publishers. New York Duvaut, G. (1976). Etude de materiaux composites elastiques a structure periodique. Homogeneisation. In Proc. Congress of Theoretical and Apllied Mechanics, Delft (1976), Koiter, ed. North-Holland. Amsterdam Duvaut, G. and Lions, J.L. (1972). Les Inequations en Mecanique et en Physique. Dunod. Paris Edwards, R.E. (1979). Fourier Series|A Modern Introduction, Vol. 1. SpringerVerlag. New York, Vol. 2. Springer-Verlag. New York, 1982 Ekeland, I. and Temam, R. (1976). Convex Analysis and Variational Problems. North-Holland. Amsterdam Fedorenko, R.P. (1978). Approximate Solution of Optimal Control Problems. Nauka. Moscow (in Russian) Fernandez Cara, E. (1989). Optimal design in uid mechanics. In Control of Partial Di erential Equations. Proc. of IFIP Conference in Santiago de Compostela

Bibliography

513

(1987), A. Bermudez, ed. Lecture Notes in Control and Information Sciences, Vol. 114, pp. 120{131. Springer-Verlag. Berlin/New York Fichera, G. (1972). Existence Theorems in Elasticity. Springer-Verlag. Berlin Fikhtengolts, G.M. (1966). A Course of Di erential and Integral Calculus, Vol. 1. Nauka. Moscow (in Russian) Fil'shtinskii, L.A. (1964). Stresses and displacements in an elastic sheet weakened by a doubly periodic set of equal circular holes. J. Appl. Math. Mech. 28, 530{543 Fix, G.J. and Strang, G. (1973). An Analysis of the Finite Element Method. Prentice Hall. Englewood Cli s, N.J. Fletcher, C.A.I. (1988). Computational Techniques for Fluid Dynamics, Vol. 2. Springer-Verlag. Berlin Flugge, W. (1957). Statik und Dynamik der Schalen. Springer. Berlin Fucik, S. Kratochvil, A., and Necas, J. (1973). Kachanov{Galerkin method. Comment. Math. Univ. Carolinae 14, 651{659 Fursikov, A.V. (1980). On some control problems and results concerning the unique solvability of a mixed boundary value problem for the three-dimensional Navier{ Stokes and Euler systems. Soviet Math. Dokl. 21, 889{893 Fursikov, A.V. (1982). Control problems and theorems concerning the unique solvability of a mixed boundary value problem for the three-dimensional Navier{Stokes and Euler equations. Math. USSR Sbornik 43, 251{273 Fursikov, A.V. (1983a). Properties of solutions of some extremal problems connected with the Navier{Stokes system. Math. USSR Sbornik 46, 323{351 Fursikov, A.V. (1983b). Some questions of the theory of optimal control over nonlinear systems with distributed parameters. Trudy Seminara Petrovskogo, no. 1, 167{189 (in Russian) Gajewski, H., Groger, K., and Zacharias, K. (1974). Nichtlineare Operatorgleichungen und Operatordi erentialgleichungen. Akademie-Verlag. Berlin Gibiansky, L.V. and Cherkaev, A.V. (1994). Microstructures of composites of extremal rigidity and exact estimates of provided energy density. In Topics in the Mathematical Modelling of Composite Materials, Kohn, R., ed. Birkhauser. New York Girault, V. and Raviart, P.-A. (1981). Finite Element Approximations of the Navier{Stokes Equations. Lect. Notes Math., Vol. 749. Springer-Verlag. Berlin/ New York Glowinski, R., Lions, J.L., and Tr^moli^res, R. (1976). Analyse Numerique des e e

514

Bibliography

Inequations Variationnelles, Tome 1, Tome 2. Dunod. Paris Goldenblat, I.I. and Kopnov, V.A. (1968). Criteria of the Strength and Plasticity of Structural Materials. Mashinostroenie. Moscow (in Russian) Gould, S.H. (1966). Variational Methods for Eigenvalue Problems. Oxford University Press. London Grigoliuk, E.I. and Kabanov, V.V. (1978). Stability of Shells. Nauka. Moscow (in Russian) Grigoliuk, E.I, Podstrigach, Ya.S., and Burak, Ya.I. (1979). Optimization of the Heating of Plates and Shells. Naukova Dumka. Kiev (in Russian) Grigorenko, Ya. M. (1973). Isotropic and Anisotropic Laminated Shells of Revolution which Possess Variable Sti ness. Naukova Dumka. Kiev (in Russian) Gunzburger, M. (1986). Mathematical aspects of nite element methods for incompressible viscous ows. In Finite Element Theory and Applications. Proc. of ICASE Conf., Hampton, Virginia (1986), pp. 124{150. Springer-Verlag. Berlin/New York Gunzburger, M. Hou, L., and Svobodny, T. (1991). Analysis and nite element approximation of optimal control problems for the stationary Navier{Stokes equations with Dirichlet conditions. Math. Mod. Numer. Anal. 25, 711-748 Gunzburger, M., Hou, L., and Svobodny, T. (1992). Boundary velocity control of incompressible ow with an application to viscous drag reduction. SIAM J. Control Optim. 30, 167{181 Guz, A.N. and Babich, I.Yu. (1980). Three-Dimensional Theory of Stability of Bars, Plates, and Shells. Vyshcha Shkola. Kiev (in Russian) Guz, A.N., Chernyshenko, I.S., Chekhov, Val.N., Chekhov, Vik.N., and Shnerenko, K.I. (1980). Theory of Thin Shells Weakened by Holes. Naukova Dumka, Kiev (in Russian) Haftka, R.T. and Gurdal, Z. (1992). Elements of Structural Optimization (Third revised edition). Kluwer. Dordrecht/Boston/London Haslinger, J. (1993). Fictitious domain approach in optimal shape design problems. In Optimal Design and Control of Structures under Statical and Vibration Response, Advanced Tempus Course. Lecture Notes, Part 1, pp. 1{8. Banach Centre. Poland Haslinger, J. and Neittaanmaki, P. (1988). Finite Element Approximation for Optimal Shape Design. Theory and Applications. John Wiley and Sons. New York

Haug, E.J. and Arora, J.S. (1979). Applied Optimal Design. Mechnical and Stuc-

Bibliography

515

tural Systems. Jhon Wiley and Sons. New York Haug, E.J., Choi, K.K., and Komkov, V. (1986). Design Sensitivity Analysis of Structural Systems. Academic Press. New York Himmelblau, D. (1972). Applied Nonlinear Programming. McGraw-Hill. London Hlavacek, I. and Necas, J. (1970). On inequalities of Korn's type. Arch. Rat. Mech. and Anal. 36, 305{334 Hlavacek, I. and Necas, J. (1982). Optimization of the domain in elliptic unilateral boundary value problems by nite element method. R.A.I.R.O. Numerical Analysis 16, 351{371 Horoshun, L.P. (1978). Methods of theory of random functions in problems of macroscopic properties of micrononhomogeneous medium. Prikl. Mekh. 14, no. 2, 3{17 (in Russian) Io e, A. and Tikhomirov, V. (1979). Extremal Problems. North-Holland. Amsterdam Ivanov, L.A., Kotko, L.A., and Krein, S.G. (1977). Boundary value problems in variable domains. In Dif. Uravn. i Primen. Trudy Sem. Processy Optimal. Upravl., Vyp. 19 (in Russian) Kantorovich, L.V. and Akilov, G.P. (1977), Functional Analysis (2nd rev. ed.). Nauka. Moscow (in Russian) English transl.: Pergamon Press. 1982 Kartvelishvili, V.M. and Kobelev, V.V. (1984). Analytical solutions to problems of optimal reinforcement for multilayer plates made of composite materials. Mekh. Komposits. Material. 19, 571{581 (in Russian) Kato, T. (1976). Perturbation Theory for Linear Operators. Springer-Verlag. New York Kohn, R.V. (1989). Recent progress in the mathematical modelling of composite materials. In Composite Material Response: Constitutive Relations and Damage Mechnisms, Sih, G.C. et al. eds., pp. 155-177. Elsevier. London Kohn, R.V. and Strang, G. (1986). Optimal design and relaxation of variational problems. Comm. Pure Appl. Math. 39, 113{137, 139{182, 353{377 Koiter, W.T. (1966). On the nonlinear theory of thin elastic shells, Parts I, II, III. Proc. Koninkl. Nederl. Akad. Wet. B69, 1{17, 18{32, 33{54 Kolmogorov, A.N. and Fomin, S.V. (1975). Introductory Real Analysis. Dover. New York Komkov, V. (1972). Optimal Control Theory for the Damping of Vibrations of Simple Elastic Systems. Springer-Verlag. Berlin

516

Bibliography

Krasnoselskii, M.A. (1956). Topological Mathods in the Theory of Nonlinear Integrall Equations. Gostekhizdat. Moscow (in Russian) Krasnoselskii, M.A., Vainiko, G.M., Zabreiko, P.P., Rutitskii, Ya.B., and Stetsenko, B.Ya. (1969). Approximate Solution of Operator Equations. Nauka. Moscow (in Russian) Krein, S.G. (1968). The behavior of solutions of elliptic problems under variation of the domain. Studia Math. 31, 411{424 (in Russian) Krizek, M. and Litvinov, W.G. (1994). On the methods of penalty functions and Lagrange's multipliers in the absract Neumann problem. Z. Angew. Math. Mech. 74, 216{218 Krys'ko, V.A. and Pavlov, S.P. (1982). Problem of optimal control of the natural frequency of inhomogeneous shells. Soviet Appl. Mech. 18, 319{325 Kupradze, V.D., Gegelia, T.G., Bashelishvili, M.O., and Burchuladze, T.V. (1979). Three-Dimensional Problems of the Mathematical Theory of Elasticity. NorthHolland. Amsterdam Kwak, B.M. and Choi, J.H. (1987). Shape design sensitivity analysis using boundary integral equation for potential problems. In Computer Aided Optimal Design: Stuctural and Mechanical Systems. Springer-Verlag. Berlin Ladyzhenskaya, O.A. (1969). The Mathematical Theory of Viscous Incomprsessible Flow. Gordon and Breach. New York Ladyzhenskaya, O.A. and Solonnnikov, V.A. (1976). Some problems of vector analysis and generalized formulations of boundary value problems for the Navier{ Stokes equations. Zap. Nauchn. Sem. LOMI, 59, 81{116 (in Russian) English transl.: in J. Soviet Math. 10 (1978), no. 2 Ladyzhenskaya, O.A. and Uraltseva, N.N. (1973). Linear and Quasilinear Elliptic Equations. Nauka. Moscow (in Russian) Laurent, P.-J. (1972). Approximation et Optimisation. Hermann. Paris Lavrentiev, M.A. and Shabat, B.V. (1973). Methods of the Theory of Functions of a Complex Variable. Nauka. Moscow (in Russian) Lekszycki, T. (1990). Eigenvalues obtimization|New view about the old problem. Control and Cybernetics 19, no. 3{4, 189{201 Lewinski, T. (1992). Homogenizing sti nesses of plates with periodic structures. Int. J. Solids Struct. 29, 309{326 Lions, J.L. (1968). Contr^le Optimal de Systemes Gouvernes par des Equations o aux Derivees Partielles. Dunod Gauthier-Villars. Paris

Bibliography

517

Lions, J.L. (1969). Quelques Methodes de Resolution des Problemes aux Limites non Lineaires. Dunod Gauthier-Villars. Paris Lions, J.L. (1983). Contr^le des Systemes Distribues Singuliers. Gauthier-Villars. o Paris Lions, J.L. and Magenes, E. (1972). Non-Homogeneous Boundary Value Problems and Applications, Vol. 1. Springer-Verlag,.New York Litvinov, W.G. (1976). Some inverse problems for bended plates. Prikl. Mat. Mekh. 40, 682{691 (in Russian) Litvinov, W.G. (1981a). A modi cation of the Ritz method for variational equations and applications to the problems with mixed boundary conditions. Di . Uravn. 17, 519{526 (in Russian) Litvinov, W.G. (1981b). Approximation of functions by a tensor product of splines and trigonometric polynomials. Ukrain. Mat. Zh. 33, 252{257 (in Russian) Litvinov, W.G. (1982a), Motion of Nonlinear Viscous Fluid. Nauka. Moscow (in Russian) Litvinov, W.G. (1982b). Optimal control by coe cients for elliptic systems. Di . Uravn. 18, 1036{1047 (in Russian) Litvinov, W.G. (1987).Optimization in Elliptic Boundary Value Problems and Applications to Mechanics. Nauka. Moscow (in Russian) Litvinov, W.G. (1990a). Control of the shape of the domain in elliptic systems and choice of the optimal domain in the Stokes problem. Math. USSR Sbornik 67, 165{175 Litvinov, W.G. (1990b). Domain shape optimization in mechnics. Control and Cybernetics 19, no. 3{4, 203{220 Litvinov, W.G. (1994). Optimization problems on manifolds and the shape optimization of elastic solids. Control and Cybernetics 23, 495{512 Litvinov, W.G. (1995). On the optimal shape of a hydrofoil. J. Optim. Theory and Appl. 85, 325{345 Litvinov, W.G. (1996). Some models and problems for laminar and turbulent ows of viscous and nonlinear viscous uids. J. Math. and Phys. Sciences. 30, 101{157 Litvinov, W.G. and Medvedev, N.G. (1979). Some questions of the stability of shells of revolution. Mat. Fiz. 26, 101{109 (in Russian) Litvinov, W.G. and Panteleev, A.D. (1980). Optimization problems for plates of variable thickness. Mech. Solids 15, 140{146

518

Bibliography

Litvinov, W.G. and Rubezhanskii, Yu.I. (1978). Methods of the theory of duality in problems of optimal loading of thin plates. Prikl. Mekh. 14, no. 2, 73{79 (in Russian) Litvinov, W.G. and Rubezhankii, Yu.I. (1982). Problems of control by the righthand sides of elliptic systems and their application to control of the stress-strain state in shells. J. Appl. Math. Mech. 46, 256{260 Litvinov, W.G. and Rubezhankii, Yu.I. (1990). On the convergence of solutions of problems of the Timoshenko theory of shells to the solution of the Kirchho {Love theory. Matem. Fiz. Nelin. Mekh. 13, 46{51 (in Russian) Loitzansky, L.G. (1987). Mechanics of Fluid and Gas. Nauka. Moscow (in Russian) Lovisek, J. (1991). Optimization of the thickness of layers of sandwich conical shells. Appl. Math. Optim. 24, 1{33 Lurie, K.A. (1993). Applied Optimal Control. Plenum. New York Lurie, K.A. and Cherkaev, A.V. (1976). On application of Prager's theorem to the problems of optimal design of thin plates. Izvest. Akad. Nauk SSSR MTT 6, 157{159 (in Russian) Makela, M.M. and Neittaanmaki, P. (1995). Nonsmooth Optimization Analysis and Algorithms with Applications to Optimal Control. World Scienti c. Singapore/New Jersey Malmeister, A.K., Tamuzh, V.P., and Teters, G.A. (1980). Strength of Polymeric and Composite Materials. Zinatne. Riga (in Russian) Marcellini, P. (1979). Convergence of second order linear elliptic operators. Boll. Un. Mat. Ital. 16-B, 278{290 Marchenko, V.A. and Khruslov, E.Ya. (1974). Boundary Value problems in Domains with Fine-Grained Boundary. Naukova Dumka. Kiev (in Russian) Marino, A. and Spagnolo, S. P (1969). Un tipo di approssimazione dell' operatore P Di (aij Dj ) con operatori j Dj (bDj ). Ann. Scuola Norm. Sup. Pisa, Cl. Sci. ij 23, 657{673 Masson, J. (1985). Method of Functional Analysis for Application in Solid Mechnics. Elsevier. New York Maz'ja, V.G. (1985). Sobolev Spaces. Springer-Verlag. Berlin Mazur, S. (1933). Uber schwache Konvergenz in den Raumen (Lp ). Studia Math. 4, 128{133 Medvedev, N.G. (1980). Some spectral peculiarity of problems of optimization of stability of shells of variable thickness. Dokl. AN Ukrainy, Ser. A, no. 9, 59{63 (in

Bibliography

519

Russian) Medvedev, N.G. and Totskii, N.P. (1984a). On multiplicity of eigenvalues in problems of optimization of stability of cylindric shells of variable thickness. Prikl. Mekh. 20, no. 6, 113{116 (in Russian) Medvedev, N.G. and Totskii, N.P. (1984b). Optimization of cylindric shells of variable thickness at axially symmetric load. Prikl. Mekh. 20, no. 9, 53{57 (in Russian) Michlin, S.G. (1962). Multi-Dimensional Singular Integral Equations. Nauka. Moscow (in Russian) Michlin, S.G. (1970). Variational Methods in Mathematical Physics. Nauka. Moscow (in Russian) Michlin, S.G. (1973). Spectrum of the bunch of operators of the elasticity theory. Usp. Mat. Nauk 28, no. 3, 43{82 (in Russian) Michlin, S.G. and Prossdorf, S. (1980). Singulare Integraloperatoren. AkademieVerlag. Berlin Mikhailov, V. (1980). Equations aux Derivees Partielles. Mir. Moscow Minoux, M. (1989). Programmation Mathematique Theorie et Algorithmes. Dunod. Paris Mota Soares, C.A., Infante Barbosa, J., and Mota Soares, C.M. (1994). Axisymmetric thin shell structures sizing and shape optimization. Control and Cybernetics 23, 513{551 Mroz, Z. and Rozvany, G.I.N. (1975). Optimal design of structures with variable support conditions. J. Optim. Theory Appl. 15, 85{101 Muschelishvili, N.I. (1963). Some Basic Problems of the Mathematical Theory of Elasticity. Noordho . Groningen Murat, F. and Simon, J. (1976). Etude de problemes d'optimal design. In Optimization Techniques: Modeling and Optimization in the Service of Man, Part 2. Proc. of the Seventh IFIP Conf., Nice (1975), J. Cea, ed. Lect. Notes in Computer Sci., Vol. 41, pp. 54{62. Springer-Verlag. Berlin Natanson, I.P. (1974). Theory of Functions of Real Variable. Nauka. Moscow (in Russian) Necas, J. (1967). Les Methodes Directes en Theorie des Equations Elliptiques. Masson. Paris Necas, J. and Hlavacek, I. (1981). Mathematical Theory of Elastic and ElasticoPlastic Bodies|An Introduction. Elsevier. Amsterdam

520

Bibliography

Neittaanmaki, P. and Tiba, D. (1994). Optimal Control of Nonlinear Parabolic Systems. Theory, Algorithms, and Application. Marcel Dekker. New York Niordson, F.I. (1985). Shell Theory. North-Holland. Amsterdam Novozhilov, V.V. (1951). Theory of Thin Shells. Sudpromgiz. Leningrad (in Russian) Obraztsov, I.F., Vasiliev, V.V., and Bunakov, V.A. (1976). Optimal Reinforcement of Shells of Revolution Made of Composite Materials. Mashinostroenie. Moscow (in Russian) Oleinik, O.A., Iosiphian, G.A., and Shamaev, A.S. (1990). Mathematical Problems of the Theory of Strongly Nonhomogeneous Elastic Medium. Izd. MGU. Moscow (in Russian) Oleinik, O.A., Shamaev, A.S., and Iosiphian, G.A. (1992). Mathematical Problems in Elasticity and Homogenization. North-Holland. Amsterdam Olho , N. (1974). Optimal design of vibrating rectangular plates. Int. J. Solids and Structures 10, 93{109 Osipov, Yu.S. and Suetov, A.P. (1990). Existence of optimal domain for elliptic problems with Dirichlet boundary condition. Preprint, Acad. Sci. Inst. Math. Mech., Sverdlovsk Panagiotopoulos, P.D. (1985). Inequality Problems in Mechanics and Applications. Birkhauser. Boston Panteleev, A.D. (1991). A problem of design of three-layered shells of revolutions according to mechanical and radioegineering characteristics. Izv. AN SSR, Mekh. Tverd. Tela, no. 6, 112{116 (in Russian) Pelekh, B.L. (1973). Theory of Thin Shells with Finite Shear Sti ness. Naukova Dumka. Kiev (in Russian) Petukhov, L.V. (1986). Optimization problems with unknown boundaries in the theory of elasticity. Izv. AN SSSR, Prikl. Mat. Mekh. 50, 231{236 (in Russian) Pironneau, O. (1974). On the optimum design in uid mechanics. J. Fluid Mech. 64, 97{110 Pironneau, O. (1984). Optimal Shape Design for Elliptic Systems. Springer-Verlag. New York Pobedria, B.E. (1984). Mechanics of Composite Materials. Izd. MGU, Moscow (in Russian) Podstrigach, Ya.S., Burak, Ya.I., Shelepets, V.I., Budz, S.F., and Piontkovskii,

Bibliography

521

A.B. (1980). Optimization and Control in the Production of Electro-Vacuum Devices. Naukova Dumka. Kiev (in Russian) Pozdeev, A.A., Niashin, Yu.I., and Trusov, P.V. (1982), Residual Stresses|Theory and Applications, Nauka, Moscow (in Russian) Pshenichny, B.N. (1980). Convex Analysis and Extremum Problems. Nauka, Moscow (in Russian) Pshenichny, B.N. (1983). Method of Linearization. Nauka. Moscow (in Russian) Pshenichny, B.N. and Danilin, Yu.M. (1978). Numerical Methods in Extremal Problems. Mir. Moscow Rao, M. and Sokolowski, J. (1989). Shape sensitivity analysis of state constrained optimal control problems for distributed parameter systems. In Lect. Notes in Control and Information Sciences, Vol. 114, pp. 236{245. Springer-Verlag. Berlin Raitum, U.E. (1989). Problems of Optimal Control for Elliptic Equations. Zinatne. Riga (in Russian) Rempel, S. and Schulze, B.-W. (1982). Index Theory of Elliptic Boundary Problems. Akademie-Verlag. Berlin Riesz, F. and Sz.-Nagy, B. (1972). Lecons d'Analyse Fonctionnelle. Akademiai Kiado. Budapest Rodi, W. (1980). Turbulence Models and Their Application in Hydraulics. I.A.H.R. Delft Rojtberg, Ja.A. (1964). Elliptic problems with nonhomogeneous boundary conditions and the local raise of smoothness up to the boundary of generalized solutions. DAN SSSR 157, 798{801 (in Russian) Rojtberg, Ja.A. (1975). Theorems of the complete set of isomorphisms for elliptic systems in the sense of Douglis{Nirenberg. Ukrain. Mat. Zh. 27, 544{548 (in Russian) Rojtberg, Ja.A. and Sheftel, Z.G. (1979). Optimal control of systems described by general elliptic problems. Diferents. Uravn. 15, 1075{1087 (in Russian) Rozvany, G.I.N. (1989). Structural Design via Optimality Criteria. Kluwer. Dordrecht Rubezhanskii, Yu.I. (1979). Some inverse problems for cylindrical shells. Prikl. Mekh. 15, no. 9, 32{36 (in Russian) Rudin, W. (1973). Functional Analysis. McGraw-Hill. New York Rvachov, V.L. and Rvachov, V.A. (1979). Nonclassical Methods of the Approxi-

522

Bibliography

mation Theory of Boundary Value Problems. Naukova Dumka. Kiev (in Russian) Sanchez-Palencia, E. (1980). Non-Homogeneous Media and Vibration Theory. Springer-Verlag. New York Schwartz, L. (1961). Methodes Mathematiques pour les Sciences Physiques. Hermann. Paris Schwartz, L. (1966). Theorie des Distributions I, II. Hermann. Paris Schwartz, L. (1967). Analyse Mathematique. Hermann. Paris Sendeckyj, G.P. (1974). Elastic properties of composites. In Mechanics of Composite Materials, Vol. 2, Sendeckyj, G.P., ed. Academic Press. New York Serovaiskii, S.Ya. (1991). Necessary and su cient optimality conditions for a system described by a nonlinear elliptic equation. Sibir. Mat. Zh. 32, 141{150 (in Russian) Serovaiskii, S.Ya. (1993a). Optimization in a nonlinear elliptic system with control in coe cients. Mat. Zametki 54, no. 2, 85{95 (in Russian) Serovaiskii, S.Ya. (1993b). Di erentiation of an inverse function in nonnormed spaces. Funk. Anal. Prilozh., no. 4, 84{87 (in Russian) Seyranian, A.P. (1973). Elastic plates and shells of minimum weight with several bending loads. Mechanics of Solids 8, no. 5, 83{89 Seyranian, A.P. (1976). Optimal beam design with limitations on natural vibration frequency and the critical force of stability. Mechanics of Solids 11, no. 1, 133{138 Seyranian, A.P. (1977). Quasioptimal solutions of a problem on optimal design with various restrictions. Soviet Appl. Mech. 13, 544{550 Seyranian, A.P. (1991). Sensitivity Analysis of Multiple Eigenvalues. Report no. 431, The Technical University of Denmark Shablii, O.N. and Medynskii, Ya.R. (1981). To the problem of creation of necessary residual stresses in a circular disk. Prikl. Mekh. 17, no. 6, 90{93 (in Russian) Shevchenko, Yu.N. (1970). Thermoplasticity under Variable Loading. Naukova Dumka. Kiev (in Russian) Simon, J. (1980). Di erentiation with respect to the domain in boundary value problems. Numer. Func. Anal. and Optimiz. 2, 649{687 Simon, J. (1989). Diferenciacion de problemas de contorno respecto del dominio. Lectures in the University of Sevilla Simon, J. (1990). Domain variation for drag in Stokes ow. In Proc. IFIP Conf. in

Bibliography

523

Shangha , Li Xunjing, ed. Lect. Notes in Control and Iformation Sciences. SpringerVerlag. Berlin Smirnov, V.I. (1959). A Course of Higher Mathematics, Vol. 5. Gos. Izd. Fiz. Mat. Lit. Moscow (in Russian) Sobolev, S.L. (1963). Applications of Functional Analysis in Mathematical Physics. Transl. Amer. Math. Soc. Mathematical Monograph 7 Sokolowski, J. (1987). Sensitivity analysis of control constrained optimal control problems for distributed parameter systems. SIAM J. Control and Optimization 25, 1542{1556 Sokolowski, J. (1990). Sensitivity analysis of shape optimization problems. Control and Cybernetics 19, 271{286 Sokolowski, J. and Zolesio, J.-P. (1992). Introduction to Shape Optimization. Shape Sensitivity Analysis. Springer-Verlag. New York Solonnikov, V.A. (1964). On general boundary value problems for systems elliptic in the sense of Douglis{Nirenberg, I. Izv. AN SSSR Ser. Matem. 28, 665{706 (in Russian) Solonnikov, V.A. (1966). On general boundary value problems for systems elliptic in the sense of Douglis{Nirenberg, II. Trudy MIAN SSSR 92, no. 4, 233{297 (in Russian) Sritharan, S. (1991). Dynamic programming of the Navier{Stokes equations. Systems and Control Letters 16, 299{307 Sritharan, S. (1992). An optimal control problem in exterior hydrodynamics. Proc. Royal Soc. of Edinburgh 121A, 5{32 Srubshchik, L.S. (1981). Buckling and Post-Buckling Behaviour of Shells. Izdat. Rostov Univers. Rostov-Don (in Russian) Suetov, A.P. (1994). A shape optimization algorithm for an elliptic system. Control and Cybernetics 23, 565{573 Sveshnikov, A.G. and Tikhonravov, A.V. (1989). Mathematical methods in problems of analysis and design of layered medium. Matem. Modelir. 1, no. 7, 13{38 (in Russian) Tartar, L. (1975). Problemes de controle des coe cients dans des equations aux derivees partielles. In Lect. Notes Econ. and Math. Syst., Vol. 107, pp. 420{426 Teters, G.A., Rikards, R.B., and Narysberg, B.L. (1978). Optimization of Laminate Composite Shells. Zinatne. Riga (in Russian) Thelen, A. (1989). Design of Optical Interference Coatings. McGraw-Hill. New

524

Bibliography

York Tikhonov, A.N., Tikhonravov, A.V., and Trubetskov, M.K. (1993). Second order optimization in problems of design of multilayered coverings. Zhur. Vych. Mat. i Mat. Fiz. 33, 1518{1536 (in Russian) Timoshenko, S. and Woinowsky-Krieger, S. (1959). Theory of Plates and Shells. McGraw-Hill. New York Triebel, H. (1978). Interpolation Theory, Function Spaces, Di erential Operators. Veb. Deutscher Verlag der Wissenschaften. Berlin Troitskii, V.A. and Petukhov, L.V. (1982). Shape Optimization of Elastic Solids. Nauka. Moscow (in Russian) Tsai, S.W. and Hahn, H.T. (1980). Introduction to Composite Materials. Technomic. Lancaster Vainberg, M.M. (1972). Variational Method and Method of Monotone Operators. Nauka. Moscow (in Russian) Vanin, G.A., Semeniuk, N.P., and Emelianov, R.F. (1978). Stability of Shells of Reinforced Materials. Naukova Dumka. Kiev (in Russian) Van Pho Phy, G.A. (1971a). Constructions of Reinforced Materials. Tekhnika. Kiev (in Russian) Van Pho Phy, G.A. (1971b). Theory of Reinforced Materials with Coatings. Naukova Dumka. Kiev (in Russian) Varga, R.S. (1971). Functional Analysis and Approximation Theory in Numerical Analysis. Soc. for Industrial and Applied Mathematics. Philadelphia Vigak, V.M. (1979). Optimal Control of Nonstationary Temperature Condition. Naukova Dumka. Kiev (in Russian) Vigak, V.M. (1988). Optimal Control of Temperature Stresses and Displacements. Naukova Dumka. Kiev (in Russian) Vigdergauz, S.B. (1977). On a case of the inverse problem of the two-dimensional theory of elasticity. J. Appl. Math. Mech. 41, 927{933 Vigdergauz, S.B. (1983). Inverse problem of three-dimensional elasticity. Mechanics of Solids 18, no. 2, 83{86 Vorovich, I.I. (1989). Mathematical Problems of the Nonlinear Theory of Shallow Shells. Nauka. Moscow (in Russian) Wang, C.-T. (1953). Applied Elasticity. McGraw-Hill. New York

Bibliography

525

Washizu, K. (1982). Variational Methods in Elasticity and Plasticity. Pergamon Press. Oxford/New York Wendland, W.L. (1985). On some aspects of boundary element methods for ellpitic problems. In The Mathematics of Finite Elements and Applications, V, pp. 193{ 227. Academic Press. London Wu, E.M. (1974). Phenomenological criteria of fracture of anisotropic medium. In Mechanics of Composite Materials, Vol. 2, Sendeckyj, G.P. ed. Academic Press. New York Yosida, K. (1971). Functional Analysis. Springer-Verlag. Berlin Zangwill, W.I. (1969). Nonlinear Programming|A Uni ed Approach. PrenticeHall. Englewood/Cli s Zavialov, Yu.S., Kvasov, B.I., and Miroshnichenko, V.L, (1980). Methods of Spline Functions. Nauka. Moscow (in Russian) Zhikov, V.V., Kozlov, S.M., and Oleinik, O.A. (1993). Homogenization of Di erential Operators. Nauka. Moscow (in Russian)

Index
averaging of a function, 40 ball closed, 6 open, 6 base of neighborhoods, 4, 9 bending moment, 213, 214, 271 bifurcation, 237 point of, 238 boundary of a set, 4 chain rule, 52 characterization of a minimizing element, 183 coe cients of the rst quadratic form, 261 completion of the set of real numbers, 3 component of the exural strain, 261 of the tangential strain, 261 of the torsional strain, 261 computation of eigenvalues, 128 condition complementing boundary, 170 limit density, 103, 112 natural, 222 of ellipticity, 169 of ltration, 436 of free edge, 222 of transversality, 222 optimality, 194 necessary, 95, 136 stable, 222 supplementary, 169 526 cone, 7 constitutive equation of a nonlinear viscous uid, 435 continuity modulus, 93 continuity of the spectrum, 118 control discontinuous, 97 nonregular, 86 nonsmooth, 97 regular, 88 convergence, 5 strong, 9 weak, 9 curvilinear coordinates, 264 cylindrical sti ness of a plate, 213 derivative Frechet, 52 G^teaux, 54 a distance, 5, 7 distribution, 38 domain in Rn , 36 domain shape optimization, 174 eigenfunction, 20 approximate, 128 eigenspace, 20 eigenvalue, 20 approximate, 128 its multiplicity, 20 element of the best approximation, 113 energy stored, 225 strain, 213

INDEX nite-dimensional approximation, 123 nite shear model of a shell, 277 form bilinear, 12 coercive, 12 symmetric, 12 free oscillations of a shell, 272 funcion continuous, 10 generalized, 38 implicit, 53 lower semicontinuous, 11 trace of, 44 test, 38 upper semicontinuous, 11 fundamental spline, 303 homeomorphism, 11 image, 10 index of elliptic operator, 174 inequality Korn, 48 Schwartz, 18 variational, 204 in mum, 3 inner point lemma, 185 interior of a set, 4 isomorphism, 11 bounded, 8 continuous, 10 convex, 7 Frechet di erentiable, 52 continuously, 52 injective, 2 k-linear, 12 sequentially continuous, 11 surjective, 2 uniformly continuous, 14 maximum, 2 maximum function, 15, 16 discrete, 15 midsurface of a shell, 260 minimum, 3 local, 71, 308 minorant, 2 multi-index, 37 neighbourhood of a point, 4 norm, 7 equivalent, 8 Euclidean, 8 observation, 194 operator coercive, 81 compact, 20 nonlinear compact, 236 selfadjoint, 20 orthogonal complement, 19 elements, 19 subspaces, 19 orthotropic material, 262

527

G-closedness of linear operators, 73 G-convergence of linear operators, 73

Kirchho hypotheses, 211, 261 Kroneker delta, 19 Lagrange principle, 71 limit lower, 3 upper, 3 majorant, 2 mapping, 2 bijective, 2 bilinear, 12

plate isotropic, 214 orthotropic, 213, 216 three-layered, 244 free oscillations of, 247 natural oscillations of, 258 problem combined, 143, 158

528 control basic, 154 general, 150 eigenvalue, 118 eigevalue optimizaton, 163 nite-dimensional, 103, 106 regular, 113 minimax control, 201 of bending of a plate, 217 of free oscillations of a plate, 223 spectrtal, 118 product of sets, 2 of topologies, 5 prototype, 10 regularization of a function, 40 Riesz method, 112 operator, 112

INDEX of revolution, 263 shallow, 279 space Banach, 7 dual, 8 Hausdor , 4 Hilbert, 18 metric, 5 complete, 6 sequentially weakly complete, 10 Sobolev, 39 topological, 4 compact, 13 metrizable, 6 separable, 4 sequentially compact, 13 vector, 6 normed, 6 strain component, 212, 213 stress component, 212 subspace of functions with zero-point strain energy, 267 supremum, 3 system ellpitic in the sense of Douglis{ Nirenberg, 168 of operators coercive, 45 W -coercive, 45 theorem Calderon, 40 imbedding, 40 Lax{Milgram, 18 Lebesgue, 38 on a composite function, 52 on equivalent norms, 41 on the invariance of Sobolev spaces, 77 on trace space, 44 Riesz, 18 topology, 4 induced, 5 weak, 9

scalar product, 18 scale of Hilbert spaces, 35 segment, 7 sequence Cauchy, 6 fundamental, 6 weakly fundamental, 10 set bounded, 2 bounded above, 2 bounded below, 2 closed, 4 clsure of, 4 compact, 13 convex, 7 dense, 4 open, 6 relatively copmact, 13 sequentially weakly closed, 10, 110 sequentially -weakly closed, 102 weakly closed, 10 shell laminated, 285

INDEX torque, 213, 214, 271

529